A Review of the Application of the Laser-Light Backscattering Imaging Technique to Agricultural Products

Abstract

Growing concerns about food safety and waste have increased consumer demand for high-quality agricultural products, particularly at the postharvest stage. This demand has prompted the development of non-destructive methods to assess or inspect the internal quality of fruits and vegetables. The backscattering imaging technique, also known as diffuse reflectance imaging, is considered a highly promising approach. Numerous studies have focused on practical applications, using laser light at selected wavelengths to develop quick multispectral methods. Due to the rapid interaction of photons with biological tissue, together with the highly computational performance of machine vision, backscattering imaging can offer a valuable alternative to traditional methods. Its primary benefits include quick measurements without chemical sample preparation, easy integration with high-throughput automatic quality control, and reduced waste, since this non-destructive technique does not damage samples. This review presents a comprehensive overview of backscattering imaging, including the measurement geometry, data analysis, and design considerations for vision systems. Additionally, it explores this technique’s advantages, challenges, and accuracy, as demonstrated using various case studies.

1. Introduction

Over the past few decades, international trade in agricultural and food products has grown significantly [1]. This growth has led many countries to establish strict quality standards for imported goods. Furthermore, as consumer expectations for product quality continue to rise, manufacturers are required to thoroughly assess their products in the production process [2]. Product quality, which is determined by attributes influencing consumer acceptance or rejection, includes factors such as nutritional content, texture, and defects [3]. In quality assessment, numerous techniques are available for evaluating product attributes. However, most of these methods are invasive and destructive, resulting in unnecessary waste, and they are time-consuming. To address these challenges, nondestructive techniques, particularly optical methods, have been widely applied to speed up the quality assessment process and to minimize waste [4].

Optical methods, including spectroscopy, machine vision, and imaging spectroscopy, that are used to evaluate the optical properties of food products provide insights into internal quality parameters, such as the SSC (soluble solids content), TA (titratable acidity), and firmness [5]. By integrating sensors, mathematical models, and advanced processing algorithms, these techniques can accurately monitor quality changes based on correlations between food quality attributes and their physicochemical properties [4]. Among the optical techniques, the use of near-infrared (NIR) spectroscopy in the assessment of the internal quality of products has seen significant development [2]. However, correlating NIR spectra with food quality attributes requires chemometric analysis, and the Beer–Lambert law may not fully describe light attenuation in highly scattering media [6]. Hyperspectral imaging is another promising technique for food quality evaluation due to its ability to characterize chemical properties in the spatial domain. Nevertheless, further development is needed for it to become the preferred non-destructive tool in this field due to its high cost and complexity [7].

Current research increasingly focuses on machine vision techniques that explore the invisible regions of the spectrum [8], with laser-light backscattering imaging (LLBI) garnering particular attention. The principle of this technique is based on light backscattering and image processing in the visible-light spectrum and short-wave near-infrared region [9]. Although photons exist across a broad spectrum, the 400–1700 nm range is commonly used in LLBI due to its cost-effectiveness and relevance to key food quality attributes [10]. This paper provides an overview of the LLBI as a developing technique, covering its theoretical and technical basics. Furthermore, the applications and recent achievements of the LLBI technique are presented, along with its limitations and recommendations for future research.

2. Concept of Backscattering Imaging

Biological materials can be classified as transparent, translucent, or opaque based on their optical properties. Agricultural products are considered to be translucent, as they allow certain wavelengths of light to pass through. When light interacts with a turbid biological tissue, it may be reflected, absorbed, or transmitted [11]. Reflectance is a complex phenomenon that occurs in three forms: regular reflectance, external diffuse reflectance, and scattering, as illustrated in Figure 1.

Regular reflectance occurs when light interacts with a mirror-like surface, such as a waxy peel. The principle of equal angles of incidence and reflection is generally followed, but reflected rays do not always return directly to the source, especially at larger angles of incidence. This phenomenon is described by the Fresnel equations (Equations (1) and (2)), illustrating the dependence of reflection on the angle of incidence (${\mathsf{\theta }}_{\mathrm{i}}$), the angle of transmission (${\mathsf{\theta }}_{\mathrm{t}}),$ and the refractive indices (${\mathrm{n}}_1$ and ${\mathrm{n}}_2$) of the media [12]. Depending on the plane of the incident electromagnetic wave, reflections are categorized as s-polarized (R_s) or p-polarized (R_p):

$${\mathrm{R}}_{\mathrm{s}}={\displaystyle \frac{{\mathrm{n}}_1\cos {\mathsf{\theta }}_{\mathrm{i}}-{\mathrm{n}}_2\cos {\mathsf{\theta }}_{\mathrm{t}}}{{\mathrm{n}}_1\cos {\mathsf{\theta }}_{\mathrm{i}}+{\mathrm{n}}_2\cos {\mathsf{\theta }}_{\mathrm{t}}}}$$

(1)

$${\mathrm{R}}_{\mathrm{p}}={\displaystyle \frac{{\mathrm{n}}_2\cos {\mathsf{\theta }}_{\mathrm{i}}-{\mathrm{n}}_1\cos {\mathsf{\theta }}_{\mathrm{t}}}{{\mathrm{n}}_2\cos {\mathsf{\theta }}_{\mathrm{i}}+{\mathrm{n}}_1\cos {\mathsf{\theta }}_{\mathrm{t}}}}$$

(2)

In contrast to regular reflection, when a light beam strikes the surface of an object, it can undergo external diffuse reflection in various directions due to surface irregularities. Light reflected at an angle of 45° is particularly informative, providing insights into the surface color [11]. Typically, only a minor fraction of the incident light (around 4–5%) is reflected diffusely at this angle, while most of the light passes through the surface and penetrates the porous tissue matrix [13]. The transmitted light undergoes further reflection and scattering within biological tissues as it interacts with internal components. This phenomenon conveys valuable information about the tissue morphology, structure, and mechanical properties. The properties of the object, such as cell size and tissue connectivity, affect the degree of light scattering. Cell wall interfaces often cause backscattering, leading to abrupt changes in reflection coefficients [14]. In addition, particles also cause scattering due to the refraction at surfaces such as those of chloroplasts and mitochondria [15]. The tissue absorbs some of the incident light, while the remaining light is either scattered or transmitted. The absorption of light by structural components in agricultural products varies with the wavelength and is determined by the structural components and the length of the light path. As a result, the characteristics of biological tissues can be described by the unique absorption and scattering properties of each material [11].

The LLBI system comprises two main components: a light source and an imaging unit. The light source emits a continuous light beam, while the imaging unit captures high-quality surface scans. Proper selection of the laser beam size and the incident angle between the sample and the beam is essential for obtaining distortion-free images. The choice of an appropriate laser wavelength is crucial for precise measurements, with wavelengths from the visible to the near-infrared range (400–1000 nm) being recommended for the non-invasive quality evaluation of food products [16]. Wavelengths in the range of approximately 620 nm to 1010 nm have potential for assessing SSC and texture characteristics, whereas the band at around 900 nm is well suited for determining moisture content. The laser beam size should be carefully selected. A larger beam diameter provides good signal quality but may impact backscattering properties, whereas smaller beam diameter concentrates the light but reduces the signal intensity, leading to noise and a smaller diffusively illuminated area (halo) [17].

The LLBI technique is a comprehensive process involving several key steps for successful implementation. These steps include selecting suitable wavelengths, preprocessing images, segmenting regions of interest, extracting relevant information and features, selecting superior features, and identifying the most suitable model for predicting the quality parameters. The thoroughness of the technique ensures careful consideration of each aspect, leading to an effective and reliable process.

2.1. Wavelength Selection

The selection of specific wavelengths is important in light backscattering imaging, as only certain wavelengths provide sufficient information about the internal quality of food. Typically, wavelengths between 400 and 1100 nm in the visible (VIS) and short-wave near-infrared (SWNIR) ranges are selected for LLBI. This bandwidth restriction is due to limitations of the silicon-based camera detectors used in these systems. Additionally, pigments of fruits and vegetables absorb light in the visible range. There are several methods for wavelength selection. The wavelength search method involves acquiring scattering images at every adjustable wavelength within the spectral range of interest. Although this approach provides comprehensive information, it is both costly and time consuming [18]. Alternatively, specific wavelengths can be selected based on their association with the quality parameters and object characteristics. For example, a wavelength of 680 nm corresponds to chlorophyll absorption in fruits, serving as a predictive indicator of fruit maturity [19]. The NIR spectroscopy results are considered more effective for wavelength selection. NIR spectroscopy provides information about the absorption and scattering characteristics of food components at different wavelengths. Consequently, wavelengths are chosen based on a strong correlation with the actual quality parameters and NIR predicted values [18]. Data mining methods, such as support vector machine (SVM) and sensitivity analysis, can be used to select wavelengths based on their ability to discriminate between quality grades or to predict specific quality attributes [20]. Principal component analysis (PCA) and partial least-squares regression (PLSR) can be used to analyze spectral data and to identify wavelengths with the highest loadings or scores, indicating their significance in explaining variations in quality parameters [21,22]. Furthermore, preprocessing techniques, such as smoothing, baseline correction, and normalization, are used to reduce noise and bias, thereby improving the model performance. These techniques also influence the spectrum and, consequently, affect the choice of wavelengths [23].

2.2. Improvement of Signal-to-Noise Ratio

Since large solid particles, vacuoles within the homogeneous tissue, and surface structures such as wrinkles and hair can affect the diffuse reflectance signal, noise reduction is important. The signal-to-noise ratio is typically improved by using a dark chamber during measurements and by increasing the hardware resolution to 12 bits per pixel [18]. Preprocessing the acquired signal can further improve its quality, with radial averaging being the most common approach. In this method, the backscattering area is segmented into circular rings around the incident point, and the average intensity within each ring is calculated. An important step in this technique is determining the incident point, which is the geometric center of the diffusively illuminated area and which corresponds to the point of peak light intensity.

According to Lu [24], the center point is defined by calculating the average pixel intensity within a predetermined window size in the central image area. This calculation is performed at a wavelength that prevents intensity saturation in any sample and has a precision of ±10 pixels. Later, Peng and Lu [25] applied a weighted center-of-gravity method to enhance the accuracy of locating incident point in scattering images. This approach successfully identified the image center across all wavelengths, even in cases of sensor saturation, outperforming previous methods. Qin and Lu [26] introduced an alternative method that leverages the symmetry of the scattering profile around the incident point. By calculating the average value, their method simplifies the subsequent analysis by focusing on only one half of the profile. In a subsequent study, image analysis using spectral averaging increased spectral resolution (nm/pixel) threefold compared with the previous method without averaging [27]. Lu and Peng [28] suggested a method for reducing the size of the image by combining information from adjacent pixels. Their technique averages pixel values over a mask (a window size of 2 $\times$ 2 pixels or larger), effectively reducing the number of pixels while preserving the overall image information.

2.3. Light-Scattering Distortion

In backscattering imaging, a complex curved surface introduces distortion in the scattered light, leading to an underestimation of the reflectance captured by the imaging system. Smaller or more curved fruits exhibit more distortions in their backscattering signals. Accurate reflectance estimation requires adjustment of the incoming light to compensate for these distortions [29]. The distortion manifests in two primary forms: intensity distortion and scattering distance distortion.

Intensity distortion in backscattering images can be corrected using the Lambertian cosine law (Figure 2), which describes how the intensity of diffusively reflected light varies with the angle of observation. To calculate the corrected reflectance R(x), the Lambertian cosine law formula (Equation (3)) is applied [30]:

$$\mathrm{R}\left(\mathrm{x}\right)={\displaystyle \frac{{\mathrm{R}}_{\mathrm{m}}\left(\mathrm{x}\right)}{\cos \mathsf{\theta }}}={\displaystyle \frac{{\mathrm{R}}_{\mathrm{m}}\left(\mathrm{x}\right)\mathrm{S}}{\sqrt{{\mathrm{S}}^2-{\mathrm{x}}^2}}}$$

(3)

where R_m(x) represents the measured reflectance, θ denotes the angle formed between the imaging direction and the pixel location being assessed on the surface, S represents the object radius (from the center to the surface), and x corresponds to the linear gap between the measured position on the surface and the incident point.

Later, Qin and Lu [31] enhanced this correction technique by focusing on a small linear element (dS) located at a distance r from the center of illumination (Figure 3). The authors further integrated this approach across the angle (θ₂–θ₁) to achieve a more accurate distortion correction (Equation (4)).

$$\mathrm{R}\left(\mathrm{x}\right)={\displaystyle \frac{\left({\mathsf{\theta }}_0+\frac{\sin 2{\mathsf{\theta }}_0}{2}\right){\mathrm{R}}_{\mathrm{m}}\left(\mathrm{x}\right)}{\left(\frac{{\mathsf{\theta }}_2}{2}+\frac{\sin 2{\mathsf{\theta }}_2}{4}\right)-\left(\frac{{\mathsf{\theta }}_1}{2}+\frac{\sin 2{\mathsf{\theta }}_1}{4}\right)}}$$

(4)

where R(x) represents the adjusted reflectance, R_m(x) denotes the measured reflectance, ${\mathsf{\theta }}_0={\tan }^{-1}\left({\displaystyle \frac{{\mathrm{r}}_{\mathrm{z}}}{\mathrm{L}}}\right)$, ${\mathsf{\theta }}_1=\mathsf{\alpha }-\mathsf{\gamma }$, ${\mathsf{\theta }}_2=\mathsf{\alpha }+\mathsf{\beta }$, $\mathsf{\alpha }={\sin }^{-1}\left({\displaystyle \frac{\mathrm{r}}{{\mathrm{r}}_{\mathrm{a}}}}\right)$, $\mathsf{\beta }={\tan }^{-1}\left({\displaystyle \frac{{\mathrm{r}}_{\mathrm{z}}+\mathrm{r}}{\mathrm{L}+\mathrm{h}}}\right)$, $\mathsf{\gamma }={\tan }^{-1}\left({\displaystyle \frac{{\mathrm{r}}_{\mathrm{z}}-\mathrm{r}}{\mathrm{L}+\mathrm{h}}}\right)$, $\mathrm{h}={\mathrm{r}}_{\mathrm{a}}-{\left({\mathrm{r}}_{\mathrm{a}}^2-{\mathrm{r}}^2\right)}^2,$ r represents the distance to the incident point, r_a stands for the assumed circular or cylindrical radius of the fruit, r_z represents the zoom lens radius and L corresponds to the distance to the object.

2.4. Segmentation of Images

Segmentation in backscattering imaging aims to isolate the region of interest (ROI) containing the diffusively illuminated surface area, thereby accelerating computation. Thresholding is a common segmentation technique used to achieve this goal by partitioning image pixels based on their intensity values. Scattered light within the visible range can be effectively evaluated through visual inspection or computational methods, such as red–green–blue (RGB) color space analysis. However, for LLBI conducted within the short-wave near-infrared (SWNIR) spectrum, computational techniques developed for grayscaled pictures, like histogram-based threshold algorithms, are more appropriate. In the research conducted by Qing et al. [32], they used bimodal thresholding to adapt the overall threshold value for scattering images using a histogram-based technique. This approach involved calculating the second derivative of the histogram to identify the inflection point for optimal thresholding.

Baranyai and Zude [33] introduced cluster analysis for segmenting valuable ROI within backscattering images. They first determined the thresholds for segmentation by examining the backscattering images, excluding saturation-induced high-intensity values and identifying the decreasing segment that represented the ROI. To derive 1D profiles from the 2D images, they applied radial averaging, leveraging the inherent radial symmetry in the diffuse illumination of the surface of the object. Subsequently, they applied a logarithmic transformation to the 1D profile.

2.5. Statistical Models

Statistical modeling involves establishing mathematical relationships between quality parameters, like firmness, moisture content, and SSC, and the measured attributes, including descriptive parameters of the scattering signal or coefficients of the fitted curve. Linear discriminant analysis (LDA) has been successfully used to classify sweet potatoes with up to 89.6% accuracy [8] and to detect decaying oranges with 94.2% accuracy [10]. The selected features for classification or regression models can be optimized using PCA [10]. The PLSR and principal component regression (PCR) models have been effectively applied to estimate the SSC and firmness of apples [9] based on the luminance histogram acquired at five wavelengths. Nanda et al. [34] used SVM to distinguish the geographic origin of citrus fruits in Indonesia. They compared different kernel functions and achieved the best classification accuracy (96.7%) with the polynomial kernel using backscattering features at 450 nm. Reported multivariate approaches used geometric features such as the diameter, area, and perimeter of the diffusively illuminated area, coefficients of fitted models, or pattern features of gray-level co-occurrence matrices (GLCM), which are typically collected at multiple wavelengths.

2.6. Deep Learning Models

Artificial Neural Network

An ANN is designed to simulate the pattern recognition of the human brain. Additionally, it can adapt and evolve dynamically, adjusting its responses to changing circumstances. This adaptability enhances its analytical capabilities, making the process more flexible and efficient [35]. The ANN is typically built with an input layer that normalizes the values, hidden layers build the connections among input data to calculate the desired output, and an output neuron to provide the result. The behavior of the neurons in the layers is determined by the activation function (linear, sigmoid, etc.). The whole network is fed with input data, and each cycle is called an epoch. The number of epochs should be large enough to enable pattern recognition but not too large to avoid overfitting. Lu [24] determined that a multilayer perceptron (MLP) network with 10 neurons in the hidden layer and trained for 20 epochs was suitable for predicting the firmness and SSC of apples. Parameters from four wavelengths were used for firmness prediction and three wavelengths were used for the SSC. Noh and Lu [36] applied a similar ANN approach using spectral data from laser-induced fluorescence at 408 nm to predict the titratable acidity (TA) of apples throughout their shelf life.

The Convolutional Neural Network (CNN) is a type of deep feedforward neural network that is inspired by the way neurons are interconnected in the human visual cortex [37]. The fundamental hierarchical structure of a CNN consists of three main layers: the convolution layer, the pooling layer, and fully connected layers. Each of these layers has a distinct function in processing the input data and converting them into a 1D feature vector, which can be used in multivariate statistical methods [38]. The CNN offers an end-to-end modeling approach for food quality evaluation by integrating denoising, dimensionality reduction, and modeling into a single process. Unlike traditional methods, which often require prior knowledge and manual effort, CNNs can automatically identify complex structures within high-dimensional data, significantly reducing the need for human intervention [39]. Zhu et al. [40] trained their CNN model on datasets with varying numbers of vegetable images. The results demonstrated that the classification accuracy decreased as the number of images was reduced. Furthermore, the CNN showed higher accuracy (92.1%) than the Back-Propagation Neural Network (78%) and the SVM classifier (80.5%). Jahanbakhshi et al. [41] applied three CNN models with 12, 13, and 15 layers to identify and evaluate visible defects in lemons. These CNN models were compared to other classifiers, including the K-nearest neighbor method, fuzzy methods, ANNs, and decision trees. Images of sour lemons were captured and sorted into two categories: healthy and damaged. The features were extracted using local binary patterns and histogram-oriented gradients. Those models were evaluated using metrics that included accuracy, sensitivity, specificity, precision, and the F-measure. The CNN achieved high scores across all metrics, with the F-measure reaching a perfect score of 1 for the 12- and 15-layer models. The results showed that incorporating a suitable pooling layer into the CNN architecture improved the performance of the model, allowing it to achieve 100% classification accuracy. Overall, the CNN models demonstrated an average accuracy of 99.93%.

The Back-Propagation Neural Network (BPNN) is a type of multilayer feedforward neural network that uses the error back-propagation algorithm for training. This technique calculates the gradient of the loss function concerning the network weights [42]. The calculated gradient is then used by an optimization algorithm to adjust the weights, aiming to minimize the loss function. For back-propagation to work, the desired output must be known for each input value, as this is necessary to compute the gradient of the loss function [43]. Wan et al. [44] developed a machine vision system to capture images of tomatoes, followed by segmentation of the regions of interest from the full images. A BPNN method was subsequently applied to classify the maturity of both Roma and pear tomatoes, achieving an accuracy greater than 99%. Zhu et al. [45] introduced an automated approach for detecting grape leaf diseases using image analysis combined with the BPNN. The study utilized Wiener filtering based on wavelet transform to reduce noise in the images. The Otsu method was used to segment diseased areas, while morphological algorithms were applied to enhance the shape of the lesions. The BPNN-based model demonstrated the capability of efficiently inspecting and classifying five different grape leaf diseases.

3. Interaction of Photons with Biological Materials

3.1. Refractive Index

The refractive index (RI) measures how light bends and refracts when it passes through materials of different optical densities. This property determines the light speed within the medium and affects light scattering, refraction, and reflection [31]. In biomedical applications, the RI is used to model photon migration within tissues. For heterogeneous biological materials, the RI is influenced by the composition and density of the tissue constituents; thus, the volume-weighted average of the values of the tissue constituents is used to estimate its value [46]. The RI of tissues of living organisms typically falls within the range of 1.33 to 1.5, with 1.33 corresponding to the RI of water. In the calculations, an RI value of 1.4 is often applied to ray-trace photon trajectories and address various boundary problems at tissue interfaces. Agro-food products tend to have lower RIs, primarily due to their high water content [46].

3.2. Absorption Coefficient

The absorption coefficient (${\mathsf{\mu }}_{\mathrm{a}}$; cm⁻¹) quantifies the rate at which the light intensity diminishes as it traverses a substance, and it is influenced by the constituents of the tissue (Equation (5)). The Beer–Lambert law establishes a logarithmic relationship between µ_a, distance, and light intensity [47]. Pigments and chromophores primarily absorb light within the 400–700 nm wavelength range, while in the short-wavelength near-infrared (SWNIR) region, water, fats, proteins, and carbohydrates become the dominant absorbers [31].

$${\mathsf{\mu }}_{\mathrm{a}}={\displaystyle \frac{-1}{\mathrm{L}}}\ln \left({\displaystyle \frac{\mathrm{I}}{{\mathrm{I}}_0}}\right)$$

(5)

where L is the distance traveled by light through the material, I₀ denotes the initial light intensity before transmission, and I represents the intensity of light after it passes through the material. The coefficient ${\mathsf{\mu }}_{\mathrm{a}}$ is expressed in cm⁻¹, and it represents the probability of absorption per unit path length.

3.3. Scattering Coefficient

The scattering coefficient (${\mathsf{\mu }}_{\mathrm{s}}$, cm⁻¹) quantifies the extent of light scattering per unit distance within a medium, and this feature is influenced by the internal microstructure of the material, particularly at the microscopic level (Equation (6)). This phenomenon is described by the following equation [47]:

$$\mathrm{d}\varnothing \left(\mathrm{L}\right)=-{\mathsf{\mu }}_{\mathrm{s}}\varnothing \left(\mathrm{L}\right)\mathrm{dL}$$

(6)

where ∅ is the power of the incident light and d∅ is the power of the scattering light. The coefficient ${\mathsf{\mu }}_{\mathrm{s}}$ is expressed in cm⁻¹ and interpreted as the probability of the collision of photons with structural elements within the unit path length.

The transport albedo (a) is a key parameter derived from absorption and scattering (Equation (7)), and it defines how the photon energy decreases after interaction. It is calculated as follows [48]:

$$\mathrm{a}={\displaystyle \frac{{\mathsf{\mu }}_{\mathrm{s}}}{{\mathsf{\mu }}_{\mathrm{a}}+{\mathsf{\mu }}_{\mathrm{s}}}}$$

(7)

where ${\mathsf{\mu }}_{\mathrm{s}}$ is the scattering coefficient and ${\mathsf{\mu }}_{\mathrm{a}}$ is the absorption coefficient.

Scattering occurs due to variations in the refractive index caused by components like membranes, air vacuoles, or organelles, which result in the dispersion of light [49]. Within biological tissues, light can undergo multiple scattering events before being re-emitted or absorbed by the medium. Thus, the scattering of light in the visible and short-wavelength NIR regions exceeds the absorption of light. Additionally, the coefficient of light scattering reduces as the wavelength increases [49].

3.4. Reduced Scattering Coefficient

The reduced scattering coefficient (${\mathsf{\mu }}_{\mathrm{s}}^{\prime }$, cm⁻¹) describes the interaction of the scattering coefficient and the average scattering angle as photons undergo diffusion within a medium (Equation (8)). During this process, photons follow a stochastic path, taking steps of varying lengths and directions. This coefficient provides important information about the scattering properties of the medium, influenced by factors such as the size and shape of particles in the medium [31].

$${\mathsf{\mu }}_{\mathrm{s}}^{\prime }=\left(1-\mathrm{g}\right){\mathsf{\mu }}_{\mathrm{s}}$$

(8)

where $\mathrm{g}$ is the anisotropy coefficient within the range of [−1,+1]. A value of −1 indicates complete backscattering, 0 denotes isotropic scattering, and +1 represents complete forward scattering. In agro-food materials, scattering is typically forward-directed, with $\mathrm{g}$ > 0.9, resulting in a significantly lower reduced scattering coefficient [33].

4. Analysis of Backscattering Images

After the images are captured, they undergo segmentation and preprocessing to eliminate noise and any extraneous details from the raw data. The refined data is then analyzed using statistical methods. Among the statistical techniques for textural analysis, the GLCM is commonly applied. This technique utilizes the variance derived from the grayscale difference histograms [50]. Techniques for analyzing texture through transformation, including the Gabor or wavelet transform, are favored due to their ability to decompose the image in the space–frequency domain [51]. A large number of textural features are usually extracted from backscattering images, which can make it challenging to develop simple and rapid techniques for quality assessment. Therefore, feature extraction methods are required for optimization. Mollazade et al. [52] suggested that extracting texture features from two-dimensional images may improve the prediction of qualitative parameters in agricultural and food commodities.

Appropriate preprocessing is essential in image analysis, as it directly impacts the accuracy and efficiency of the results. For instance, methods like radial and profile averaging can effectively reduce the computations in feature extraction, thereby simplifying the subsequent analysis [30]. Spectral averaging can improve the signal quality. However, this process also increases the processing time, which can be a significant drawback in real-time applications [52]. Pixel binning is another technique that can help to reduce noise but often leads to blurred images. This can impact the accuracy of feature extraction in pattern analysis and curve fitting [52].

Various methods can be applied to analyze the segmented backscattering images, including optical scattering [53] and light intensity [54] and radial averaging [55]. These approaches generate a data profile that is subsequently fitted by models like the Lambert–Beer (LD), modified Lambert–Beer (MLD), and generalized Lambert–Beer (GL) models. These models quantify the relationship between light attenuation and the physical and chemical properties of the scattering medium. The LD model is presented in Equation (9).

$${\mathrm{I}}_{\mathrm{wi}}={\displaystyle \frac{{\mathrm{a}}_{\mathrm{wi}}}{1+{\left(\frac{\mathrm{x}}{{\mathrm{b}}_{\mathrm{wi}}}\right)}^2}}$$

(9)

where I represents the light intensity; x denotes the scattering distance from the center of the incident beam in millimeters; a corresponds to the peak value of the scattering profile at x = 0; b represents the full width at half maximum (FWHM) of the scattering profile measured in millimeters; and wi represents a specific wavelength, with i ranging from 1 to n, where n is the total number of wavelengths. The MLD model is presented in Equation (10).

$$\mathrm{R}={\displaystyle \frac{\mathrm{b}}{{\left(1+\frac{\mathrm{x}}{\mathrm{c}}\right)}^{\mathrm{d}}}}$$

(10)

where R represents the intensity of the backscattered light at x distance from the incident beam, b corresponds to the peak value of the profile, c denotes the FWHM, and d represents the slope around the FWHM. The GL model is presented in Equation (11).

$$\mathrm{R}\left(\mathrm{r}\right)=\mathrm{a}+{\displaystyle \frac{\mathrm{b}}{\left[1+\mathrm{e}{\left(\frac{\mathrm{r}-\mathrm{c}}{\mathrm{d}}\right)}^2\right]\exp \left[\frac{1-\mathrm{e}}{2}{\left(\frac{\mathrm{r}-\mathrm{c}}{\mathrm{d}}\right)}^2\right]}}$$

(11)

where R(r) is the light intensity at a distance r (mm) from the incident point, a is the asymptotic value of the light intensity as $\mathrm{r}$ approaches infinity, b is the peak value of the light intensity at the center, c is the intensity peak location, d is the FWHM, and e is the shape factor. The shape parameter e is defined in the range of 0–1, where 0 indicates a pure Gaussian function, and 1 represents a pure Lorentzian.

The software LightScatter (version R11) offers various techniques for analyzing backscattering images, which are classified into three categories [56]. The first category, called “spatial analysis”, uses a spatial domain approach by determining the parameters for a function fitted to the 1D profile. In horticultural products, photon scattering is typically symmetrical relative to the incident point, allowing the use of the radial averaging method to calculate the 1D profile (Figure 4). The software identifies the center of the saturated area in the image and then calculates the mean intensity from pixels situated within concentric rings. Parameters for radial averaging, such as the radius of the initial ring (saturated region), ring width, and the number of rings, can be adjusted. Spatial analysis techniques can be used by fitting various functions like Farrell’s diffusion theory model, and the modified Lorentzian, modified Gompertz, and Gaussian-Lorentzian models [30,57,58,59].

The second category of approaches for analyzing backscattering images involves extracting features based on descriptive statistics of the pixel intensity values, such as the mean, mode, median, standard deviation, coefficient of variation, skewness, and kurtosis [43]. The final category focuses on texture analysis and can be divided into statistical, transform-based, and model-based techniques. In statistical methods, texture is assessed indirectly through the analysis of the distribution and relationship between the grayscale values in an image. Transform-based techniques involve converting the image from one domain to another to extract features related to texture characteristics. Generative and stochastic models are applied for texture representation, relying on the correlation between pixel grayscale values and those of their neighboring pixels [60].

Monte Carlo Simulation

The Monte Carlo method is a numerical approach utilized to ray trace the photons as they traverse from one interaction to another. The behavior of light propagation within tissues is inherently stochastic and can be accurately replicated through computer simulations using weighted random absorption and scattering events, as described by Pavlin et al. [61]. In biomedical applications, where scattering appears as noise on medical images, the Monte Carlo technique is preferred due to its accuracy, especially in regions near light sources and boundaries. This method offers significant flexibility, allowing the simultaneous tracking of multiple physical parameters with the desired spatial and temporal resolution. However, due to its statistical nature, this technique requires substantial computational time to achieve precise simulation outcomes [62].

Multiple approaches can be used to perform Monte Carlo simulations. The most common one estimates the steady-state distribution of light. In this method, a photon is introduced into the biological medium at the incident point, and its path through the medium is determined based on the known optical properties (absorption, scattering, anisotropy) of the material being studied. This simulation tracks large number of photons, allowing the assessment of various physical quantities, such as the diffusion coefficient, penetration depth, and internal photon absorption [6].

Wang et al. [63] introduced a computer program utilizing the Monte Carlo technique to simulate light transport in multilayered tissues. It is considered the reference for Monte Carlo simulation software solutions. Since its introduction, this method has been widely adopted for addressing complex optical challenges encountered in tissue analysis. When combined with machine vision technology, it has been applied in biosystems engineering for determining the optical properties of horticultural products, facilitating quality assessment and grading through inverse modeling [33].

5. Comparison of LLBI and Other Imaging Methods

LLBI is rapidly emerging as a promising technique for the quality assessment of agricultural products [64]. Its efficiency and real-time application potential show great promise, especially in industrial environments where speed and cost are important factors. LLBI technology uses different image processing methods compared with other techniques like hyperspectral imaging (HSI), Raman imaging (RI), and magnetic resonance imaging (MRI). Therefore, it possesses unique advantages and disadvantages when compared with these methods. LLBI typically uses specific wavelengths within the range of 400–1700 nm that are associated with specific quality attributes of the sample. LLBI illuminates biological tissues with laser light and examines the scattered light. By studying this backscattered light, LLBI can reveal information about both the surface and the internal properties of the sample, such as defects and hardness [10]. In contrast, HSI collects data over a wide spectral range of 400–2500 nm. It constructs 3D data sets, known as hypercubes, by recording spectral information for each pixel in a 2D array. These hypercubes, comprising two spatial dimensions and one spectral dimension, enable HSI to effectively detect spatial structures and chemical compounds [65]. RI offers a clear signal for identifying the chemical forms and crystal locations within samples. Spectral images are obtained through techniques like point-by-point, line scanning, and area scanning. Therefore, RI is primarily used to detect the molecular composition and chemical variations within a sample [66]. MRI produces high-quality 2D and 3D images by mapping the proton molecule density, taking advantage of proton abundance and activity. This method is useful for detecting changes in the physical structure, such as aging and defects [67].

Whether the technique is LLBI, HSI, RI, or MRI, the process of band selection and feature extraction remains important for enhancing the data quality and improving model performance. Approaches for band selection such as PCA, PLSR, SVM, and ANN are commonly used in these techniques. However, since each technique operates on different principles, careful consideration is required when selecting the appropriate methods for wavelength selection. HSI, which deals with complex and high-dimensional hypercube datasets, necessitates the use of more advanced band selection techniques. In contrast, LLBI, which works with simpler data, allows for a more straightforward band selection process that primarily focuses on scattering profiles instead of the extensive spatial and spectral information captured by HSI [52,68]. Band selection in RI focuses on identifying narrow spectral bands associated with molecular vibrations, whereas LLBI utilizes a single wavelength laser to analyze the backscattered signal, providing information about the tissue microstructure and composition. This difference makes LLBI less suitable for detailed chemical composition analysis but better suited for assessing structural quality [69]. In contrast with the other techniques, MRI operates on parameters like relaxation times and proton density, rather than spectral bands. This fundamental difference means that while MRI emphasizes internal structural details, LLBI concentrates on light diffusion patterns. Consequently, band selection in LLBI involves optimizing specific wavelengths, whereas MRI involves optimizing different types of parameters [67].

The principles of these techniques indicate that LLBI generates low-dimensional data, allowing for the rapid and cost-effective assessment of agricultural product characteristics. However, this technique has limitations in performing in-depth chemical analysis, lacks the depth resolution of MRI, and is less effective at detecting complex compositional changes compared with HSI and RI. In general, techniques like HSI, RI, and MRI offer more comprehensive insights, but they also have higher complexity, longer processing times, and higher costs. This makes LLBI ideal for implementing real-time, high-throughput quality control in industrial settings.

6. Applications

Numerous studies have explored the use of LLBI for assessing and monitoring the quality of agricultural products. A summary of these studies is presented in

Table 1. A summary of LLBI applications in agricultural products.

Reference	Product	Wavelength (nm)	Calibration Model *	Results
Arefi et al. [70]	Apple slides	635, 980, and 1450	GPR and PLSR	The GPR models at 980 nm provided the most accurate predictions of quality changes in apple slices during the drying process. Predictions for moisture content achieved an R2 value of 0.92; vitamin C had an R2 of 0.79, and the SSC reached an R2 value of 0.88.
Ali et al. [71]	Watermelon	658	PLSR	Predicting the changes in firmness, SSC, pH, and moisture. The R2 for quality prediction ranged from 0.88 to 0.94.
Lockman et al. [72]	Cocoa pods	658 and 705	MLR	The prediction of cocoa fruit ripeness at 705 nm obtained an R2 value of 0.65 for hardness and 0.80 for chroma. Additionally, the 705 nm wavelength demonstrated outstanding classification performance, achieving 95% accuracy in categorizing samples based on their maturity level.
Babazadeh et al. [73]	Potato	532 and 635	ANN	Combining LLBI with the ANN method could successfully classify potato cultivars with an accuracy exceeding 90%. Moreover, the method achieved over 98% accuracy in distinguishing between healthy and toxic potatoes.
Qing et al. [9]	Apple	600–1100	PLSR and SMLR	The assessment of apple fruit development focused on the SSC and firmness. The PLSR prediction model demonstrated the highest R2 value, exceeding 0.88. Wavelengths around 780 and 880 nm were found to provide important information related to the SSC and firmness of apple.
Zulkifli et al. [17]	Banana	658	MLR	The mean intensity values and cross-sectional area were reliable for assessing the changes in the physicochemical properties of banana during ripening. The classification of unripe and ripe bananas reached an accuracy of 94.2%
Nanda et al. [34]	Citrus fruit	450, 532, and 648 nm	-	The combination of the gray-level co-occurrence matrix method and the support vector machine algorithm was used to extract texture features and to develop a classification model. The proposed approach reached an accuracy of 96.667% for authenticating the geographic origin.
Yang et al. [64]	Kiwifruit	830	LR	The prediction of kiwifruit flesh firmness during storage showed great accuracy, with R2 values greater than 0.9 for both the ”Zesy002” and ”Hayward” cultivars.
Onwude et al. [74]	Sweet potato	658	PLSR	Good correlation was obtained with the moisture content and color properties of sweet potato during drying, achieving an R2 > 0.7.
Udomkun et al. [75]	Papaya	532, 650 and 780	MLR	The model at 650 nm provided the best fit for changes in papaya during the drying process, including those affecting the moisture content, shrinkage, and color. Applying MLR based on the illuminated area and light intensity parameters provided the most accurate models for predicting all quality attributes.

* ANN: Artificial Neural Network; GPR: Gaussian process regression; LR: linear regression; MLR: multi-linear regression; NR: nonlinear regression; PLSR: partial least-squares regression; SMLR: stepwise multiple linear regression.

, which provides an overview of the findings in this field. Overall, the applications of LLBI can be categorized into three main areas: the inspection of fruit quality, the monitoring of postharvest processing, and the control of food quality throughout various stages of production and storage.

6.1. Fruit Quality Inspection

The sensory evaluation of fruits by consumers is an important factor in determining their ripeness and overall quality, which can fluctuate based on the fruit variety and ripeness stage. Over the past few decades, various optical techniques have been developed to assess both the internal and external quality indices of fruits, but these methods are often expensive and require advanced instruments (e-nose or portable spectrophotometer). LLBI offers a comparatively affordable and easily implementable alternative to analytical measurements and internal quality assessment by magnetic resonance imaging (MRI), nuclear magnetic resonance (NMR), and X-ray imaging. Although the majority of studies in this area focused on internal quality parameters, a limited number of reports are available on the application of LLBI.

The study by Mollazade and Arefi [56] investigated the use of different models to classify apples based on their freshness and mealy texture. Mealy apples exhibited an increased absorption coefficient at 650 nm and a decreased absorption coefficient at 980 nm, which was attributed to the breakdown of carbohydrates and internal browning. Mealy apples also showed a decreased scattering coefficient at 650 nm and an increased scattering coefficient at 980 nm. Different behavior was observed for apricot by the same authors [76], possibly because their assessment was limited to 650 nm, and was probably influenced by the presence of chlorophyll in the fruit. Apple classification was performed using Farrell model parameters, as well as modified Lorentzian, modified Gompertz, and Gaussian-Lorentzian models. Discriminating semi-mealy apples from fresh or mealy apples was challenging. However, by grouping fresh and semi-mealy apples as non-mealy, an ANN classification combining both 650 nm and 980 nm achieved a good classification accuracy of 74%. Tu et al. [77] found that fruit size did not significantly affect the total number of acquired photons in the scattering images, but there were significant differences (p < 0.05) between the back and blush sides of the fruit. Additionally, the area of the backscattering region from the apple illuminated by the laser beam increased significantly (p < 0.05) with fruit ripening. In a study conducted by Baranyai et al. [78], the LLBI technique was applied at a wavelength of 785 nm to monitor the optical properties of apple tissue at three different ripening stages: unripe, ripe, and overripe. The apples were stored under controlled temperature and atmospheric conditions for 160 days. Backscattering profiles were compared to Monte Carlo simulations to estimate the anisotropy factor and total interaction coefficient. The results revealed an increase in the total interaction coefficient during the initial 81 days of storage, followed by a subsequent decrease. Additionally, a steady decline in the anisotropy factor was observed throughout the storage period. These findings suggested that the anisotropy factor holds significant information and can be added as a parameter for monitoring the ripeness of apples.

So far, fruit firmness has been the primary predicted parameter. Apart from that, estimation of the SSC and TA have been explored in reports as well. However, achieving a robust calibration for these indices was more challenging than firmness prediction. Noh and Lu [36] reported a correlation coefficient of approximately 0.75 for apple firmness prediction using a PCA-ANN model. In contrast, the correlation coefficients for SSC and TA were 0.66 and 0.57, respectively. Similar results were reported for apple by Qing et al. [9].

6.2. Postharvest and Process Monitoring

Drying is an important fruit processing technique for eliminating the need for chemical preservatives and for reducing packaging and transportation costs through size and weight reduction. To ensure optimal drying results and to prevent over-drying, it is essential to continuously monitor the quality parameters throughout the drying process.

Various contemporary technologies, including MRI, machine vision, and NIR spectroscopy, have been applied to the non-invasive monitoring of the drying process. A relatively recent and promising addition is the LLBI technique, which first found an application in monitoring fruit drying processes in 2008 [9]. Studies have assessed the capability of backscattering imaging to predict changes in color, SSC, and hardness in different fruits during the drying process.

LLBI has demonstrated promise in predicting moisture content changes in fruits, particularly when drying at lower temperatures. As the drying duration extends, the turgor pressure against cell walls decreases, resulting in a rougher fruit surface and increased optical density, which, in turn, reduces the number of scattering photons emitted from the samples [79]. Furthermore, it has been observed that luminance is an effective predictor of changes in the moisture content and SSC of apples, whereas laser scattering at 635 nm was unsuitable for predicting variations in firmness [80]. Another study by Siyum et al. [81] successfully detected the effect of ascorbic acid pretreatment on the quality of banana slices during drying with LLBI. It was found that the ascorbic acid content and drying time significantly affected the extracted parameters at 532, 635, 780, and 1064 nm.

6.3. Food Quality Control

Quality assurance is a critical aspect of the food processing industry, with a primary focus on evaluating processed food items, predicting their quality, and ensuring that they adhere to established standards. The development of precise inspection methods for automated quality control systems in the food supply chain is of great importance. Brosnan and Sun [82] conducted a comprehensive review of computer-vision-based techniques applied to quality assessment in the food sector. Among these methods, one emerging approach is low-coherence laser backscattering imaging, which has shown promise in assessing the quality of dairy and meat products, albeit primarily within laboratory settings.

Qin and Lu [29] explored the potential of LLBI as a control mechanism for estimating milk fat content. They captured hyperspectral backscattering images of milk samples with varying fat levels ranging from 0% to 3.25%. The milk fat content exerted a noticeable influence on absorption and scattering coefficients, with the scattering profile increasing as the fat content decreased. Specifically, the absorption and reduced scattering coefficients at the 600 nm wavelength demonstrated strong predictive capabilities, boasting correlation coefficients of 0.995 and 0.998, respectively.

Regarding meat quality control, maintaining quality standards throughout the entire process, from cattle slaughter to meat distribution, is essential for meeting consumer expectations. Several destructive and non-destructive techniques, such as chemical procedures, NIR spectroscopy, electronic tongue, and electronic nose, have been applied to provide insights into meat quality. Tenderness is a key factor influencing meat quality within the beef industry, directly impacting price and consumer repurchase decisions. Cluff et al. [53] investigated the feasibility of LLBI as a non-destructive means for predicting beef steak tenderness. They acquired hyperspectral backscattering images across the wavelength range from 496 nm to 1036 nm. By utilizing stepwise regression, they successfully estimated beef steak tenderness based on the Warner–Bratzler shear force (WBS) and parameters associated with the modified Lorentzian distribution function. Notably, specific wavelengths (501, 510, 646, 651, 927, 1005, and 1023 nm) were of greater importance in constructing the calibration model. Their system achieved a correlation coefficient of 0.67 for WBS value predictions, indicating the potential for further research to enhance the accuracy of LLBI for practical use.

Lorente et al. [59] applied five laser diodes spanning the visible to NIR spectra and captured images of fruit samples, both with and without fungal inoculation. Subsequently, they analyzed the backscattering profiles using the Gaussian–Lorentzian function. For classification purposes, they applied LDA and achieved an accuracy of 80.4% at 532 nm. Notably, combining information from five different wavelengths (532, 660, 785, 830, and 1060 nm) significantly improved the accuracy to up to 96.1%. This study showed the efficiency of LLBI, where the selected function fits well for an acquired signal with an average R² = 0.998.

7. Future Prospects

Despite remarkable advancements in the LLBI technique over the past decade, several challenges remain to be addressed in this field. Numerous studies have demonstrated the effectiveness of backscattering imaging in evaluating the internal properties of agricultural and food products, such as the moisture content, SSC, and firmness. However, internal defect detection is lacking, which is also considered an important issue in the assessment of fruit and vegetable quality. Thus, applying the LLBI technique in the quality inspection of fruit and plant-based materials holds significant promise in this field. Another major challenge is the ability to achieve continuous and real-time evaluation, primarily due to limitations associated with the imaging technology used in these systems. The application of LLBI in automatic sorting lines requires surface scanning instead of sampling from a small surface area. This leads toward line laser modules instead of focused beams. To enhance the processing speed and image analysis efficiency, there is a need for further advancements in the field of image processing algorithms. Furthermore, it is important to search for the most suitable wavelengths for specific applications. Mathematical models also need alternative solutions to nonlinear regression, which are slow to deploy in high-capacity grading. Moreover, to thoroughly evaluate the effectiveness of this technique, it is suggested to extend the research with more product categories beyond fruits and vegetables, such as transparent liquids (e.g., wine) and complex optical structures (e.g., egg) [2,83].

8. Conclusions

This review evaluated the current state of light backscattering imaging for the quality evaluation of agro-food products, providing an in-depth understanding of the fundamentals of this approach. It highlighted key aspects, including the choice of laser-light wavelengths, the parameters used for prediction, and the methodologies applied to construct calibration and prediction models. Significant results have been achieved using backscattering imaging to evaluate the internal characteristics of agricultural and food products. However, challenges remain for the efficient deployment of this system on an industrial scale. It is notable that most published studies focused on batch assessments, which do not accurately simulate real-time online sorting and grading situations in terms of the rapid evaluation of a high volume of products. Despite these limitations, ongoing research efforts aim to address these hurdles and to create cost-effective, high-performance laser-based equipment, which is expected to drive the adoption of this technology. Looking forward, promising results from studies indicate the effectiveness of this technique as an innovative tool for quality assessment.