A New Deep Learning Model for the Classification of Poisonous and Edible Mushrooms Based on Improved AlexNet Convolutional Neural Network

Abstract

The difficulty involved in distinguishing between edible and poisonous mushrooms stems from their similar appearances. In this study, we attempted to classify five common species of poisonous and edible mushrooms found in Thailand, Inocybe rimosa, Amanita phalloides, Amanita citrina, Russula delica, and Phaeogyroporus portentosus, using the convolutional neural network (CNN) and region convolutional neural network (R-CNN). This study was motivated by the yearly death toll from eating poisonous mushrooms in Thailand. In this research, a method for the classification of edible and poisonous mushrooms was proposed and the testing time and accuracy of three pretrained models, AlexNet, ResNet-50, and GoogLeNet, were compared. The proposed model was found to reduce the duration required for training and testing while retaining a high level of accuracy. In the mushroom classification experiments using CNN and R-CNN, the proposed model demonstrated accuracy levels of 98.50% and 95.50%, respectively.

1. Introduction

Thailand is a biodiverse country with a vast array of flora and fauna. Fungi, commonly known as mushrooms, make up a significant portion of these species and play a role in the lifecycles of many living organisms. Moreover, mushrooms are highly nutritious, a good source of protein, low in calories and unsaturated fats, and a rich source of vitamins and iron [1]. The most popular species are eaten as vegetables in several meals. Mushrooms are also economically important, as wild mushrooms are gathered and sold. Mushrooms can be found after 2–3 days of rain between May and September. Mushrooms appear in large numbers in the dipterocarp forests, mixed forests, and dry evergreen forests in the north and northeastern parts of Thailand. There are between two and three million different species of mushrooms in the world, and these can be divided into two types: poisonous and edible [2,3,4]. Certain poisonous mushrooms have very similar appearances to edible mushrooms. According to folk wisdom, classifying mushrooms as poisonous or edible involves boiling mushrooms in the same pot as rice: a change in the color of the rice demonstrates that the mushrooms are poisonous. A silver spoon changing from silver to black when used to stir a pot of boiling mushrooms is another folk method used to identify poisonous mushrooms. However, the above experiments are not precise, as certain poisonous mushrooms do not respond to these tests. The basic morphological characteristics of poisonous mushrooms are the presence of bright, colorful scales on the cap and a circle under the cap. Inexperienced gatherers can easily make a mistake. Poisonous mushrooms affect the nervous system, leading to death if consumed in high quantities [5]. Every year, people die from eating poisonous mushrooms in Thailand.

Deep learning, which is part of the machine learning (ML) concept, automates learning by mimicking the operation of the human brain. It involves the implementation of multiple layers of neural networks, allows for complex processing, and is highly accurate. Additionally, deep learning methods are more accurate than human classifiers because they are trained by large datasets. A model can be trained using sample data, and its accuracy depends on the amount of sample data used. Presently, deep learning is used in various fields, such as facial recognition [6,7], plant disease detection [8,9], and autonomous vehicles, for object classification to increase the speed of work [10,11]. Deep learning consists of three components: the input layer, the hidden layer, and the output layer.

The proposed model was used to improve the AlexNet convolutional neural network in order to increase the speed and maintain the accuracy when classifying poisonous and edible mushrooms. The authors hope that this may reduce the number of deaths caused by the consumption of poisonous mushrooms. The results of the proposed model were compared with those of three pretrained architectures, AlexNet, ResNet-50, and GoogLeNet, for the classification of five species of poisonous and edible mushrooms: Inocybe rimosa, Amanita phalloides, Amanita citrina, Russula delica, and Phaeogyroporus portentosus.

2. Related Work

2.1. Convolutional Neural Network (CNN)

A convolutional neural network, a multiple-layer neural network model, is a popular deep learning algorithm that is used for image analysis and object recognition [12]. The CNN contains convolutional layers, pooling layers, a rectified linear unit, fully connected layers, and a softmax layer [13,14,15]. The structure of the CNN is shown in Figure 1.

The convolution layer uses filters to extract features from an image [16]. The filters are smaller than the input because they scroll through each portion of the image to detect features. An example using a 3 × 3 filter is shown in Figure 2.

The rectified linear unit (ReLU) is a nonlinear activation function. This layer has a value of zero if the input is negative, but it will return the value x when the input is positive [17]. ReLU is defined by the following equation:

$$f\left(x\right)=\max \left(0,x\right)$$

(1)

The pooling layer reduces the number of parameters and, depending on the pool size, minimizes the complexity of the output and reduces the training time [12,18,19]. There are different types of pooling: max pooling, average pooling, global average pooling, global max pooling, mixed pooling, L_p pooling, stochastic pooling, spatial pyramid pooling, and region of interest pooling [20,21,22,23]. An example of max pooling, with a pooling size of 2 × 2, is shown in Figure 3.

The flattened layer converts the multidimensional data to one dimension before forwarding them to a fully connected layer.

The fully connected layer connects all of the neurons by combining all properties extracted from the previous layer for use in classifying the following layer [16,24].

The softmax layer is a functional layer that takes the input and returns the probability of classifying objects [17].

2.2. Region Convolutional Neural Network (R-CNN)

A CNN is designed to classify objects; however, it does not detect objects or draw bounding boxes around objects. Therefore, the R-CNN method was proposed for object detection and to draw bounding boxes around objects [25]. In addition to the R-CNN method, there are many other models, such as Fast-RCNN and Faster-RCNN. In this research, the R-CNN method was used because it is a basic method that can detect objects without complexity, and it can easily be modified. The R-CNN consists of four operational steps: first, a selective search of the image is used to extract 2000 region proposals, and each proposal is reshaped to the same size and passed on as an input to a CNN. Second, the CNN uses feature extraction for each region proposal [26]. Third, upon obtaining the extractions, these features are used to classify region proposals using the SVM. Finally, bounding boxes are created around objects. The R-CNN architecture is shown in Figure 4.

2.3. AlexNet Architecture

AlexNet is a CNN with eight layers; the architecture has five convolution layers, three max pooling layers, and three fully connected layers [27,28]. AlexNet was trained on more than one million images and over 1000 categories from the ImageNet database. It has an image input size of 227 × 227 × 3 pixels: 227 refers to the width and height of the input images and 3 distinguishes the images as RGB color images. The first convolution layer consists of 96 filters with a filter size of 11 × 11 and four strides. The second convolution layer consists of 256 filters with a filter size of 5 × 5 and one stride. The third convolution layer consists of 384 filters with a filter size of 3 × 3 and one stride. The fourth convolution layer consists of 384 filters with a filter size of 3 × 3 and one stride. The fifth convolution layer consists of 256 filters with a filter size of 3 × 3 and one stride. The convolution layer is shown in

Table 1. AlexNet Architecture Convolution Layer.

Layer	Filters	Filter Size	Strides
1	96	11 × 11	4
2	256	5 × 5	1
3	384	3 × 3	1
4	384	3 × 3	1
5	256	3 × 3	1

. Afterwards, each convolutional layer is normalized using ReLU and max pooling with a pool size of 3 × 3 [29,30]. The AlexNet architecture is shown in Figure 5.

2.4. ResNet-50 Architecture

ResNet-50 is a CNN that is fifty layers deep. The architecture has an image input size of 224 × 224 × 3 pixels, 16 bottleneck building blocks, 48 convolution layers, and 1 fully connected layer [31,32]. It contains both the same and different bottleneck building blocks, as shown in Figure 6. Bottleneck building blocks numbers 1 to 3 contain convolution layers with 64 filters with a filter size of 1 × 1, another layer with 64 filters with a filter size of 3 × 3, and 256 filters with a filter size of 1 × 1. Building block numbers 4 to 7 contain two convolution layers with 128 filters with filter sizes of 1 × 1 and 3 × 3, respectively, and a convolution layer with 512 filters with a filter size of 1 × 1. Building block numbers 8 to 13 contain two convolution layers with 256 filters with filter sizes of 1 × 1 and 3 × 3, respectively, and another layer with 1024 filters with a filter size of 1 × 1. Building block numbers 14 to 16 contain two convolution layers with 512 filters with filter sizes of 1 × 1 and 3 × 3, respectively, and another layer with 2048 filters with a filter size of 1 × 1 [33]. There are various ResNet models, e.g., ResNet-18, ResNet-50, and ResNet-101. The ResNet-50 architecture is shown in Figure 7.

2.5. GoogLeNet Architecture

GoogLeNet is a CNN that is 22 layers deep [34]. It has an image input size of 224 × 224 × 3 pixels and was designed to incorporate the concept of an inception module [35]. The inception module contains a convolutional filter with a filter size of 1 × 1, a convolutional filter with a filter size of 3 × 3, a convolutional filter with a filter size of 5 × 5, and a 3 × 3 max pooling pool [36,37,38]. GoogLeNet was designed to have nine inception modules, as shown in Figure 8. When data are sent to a module, they are divided into four groups and merged into one set when leaving the module. The inception module is shown in Figure 9.

3. Experimental Section

This section discusses the architecture used for learning and presents the mushroom dataset. The study was performed on MATLAB R2021b, using Windows 10 64-bit, on a CPU with Intel core i5-12600, GPU NVIDIA GeForce RTX 3060, and 12 GB of Ram.

3.1. Proposed Model

Learning technique transfer involves taking a model that has been pretrained for one task and adapting it to the new task so that the model does not have to be retrained entirely. This reduces the training time. The proposed model was developed to transfer learning and improve AlexNet by removing the fourth and fifth convolution layers from the original AlexNet Model and adding the GoogLeNet inception module layer. The proposed model consists of an image input size of 227 × 227 × 3 pixels, three convolution layers, one inception module, and three fully connected layers. The first convolution layer consists of 96 filters with a filter size of 11 × 11, ReLU, and a max pooling size of 3 × 3. The second convolution layer consists of 256 filters with a filter size of 5 × 5, ReLU, and a max pooling size of 3 × 3. The third layer is the inception module, as shown in Figure 9. The fourth convolution layer consists of 384 filters with a filter size of 3 × 3, ReLU, and a max pooling size of 3 × 3. The proposed architecture is shown in Figure 10.

3.2. Mushroom Dataset

The mushroom dataset used in this research included 5 species, 623 images, and an image size of 227 × 227 × 3 pixels. It was divided into two parts. The mushroom data were composed of poisonous and edible mushrooms including the edible types Amanita citrina (A), Russula delica (B), and Phaeogyroporus portentosus (C) and the poisonous types Inocybe rimosa (D) and Amanita phalloides (E). The data are presented in

Table 2. Mushroom dataset.

Edible Mushrooms			Poisonous Mushrooms
A	B	C	D	E
248	88	155	76	56

and shown in Figure 11.

Deep learning is a highly accurate method because it uses a large dataset for training; however, if the size of the dataset used in the training is insufficient, unreliable answers can be produced. Therefore, data augmentation is used to increase the amount of data used in training in order to avoid overfitting and ensure the model’s accuracy [39].

From the mushroom dataset of 623 images, the authors generated a dataset containing 2000 images and divided it into 10 equal groups; these were used in training and testing based on the K-fold cross-validation method, as shown in

Table 3. K-Folds cross-validation.

Group	Amount	Group	Amount
1	200	6	200
2	200	7	200
3	200	8	200
4	200	9	200
5	200	10	200

. The dataset was divided into training and test sets using a ratio of 90:10 [40].

3.3. Confusion Matrix

Before implementing a model, it is necessary to measure its performance to determine whether it is effective enough to be developed or used. A confusion matrix is a tool that describes the performance results for the predictions made by models built with machine learning. The authors measured accuracy using Equation (3), precision using Equation (4), recall using Equation (5), and the F1 score using Equation (2) as the standard metrics to evaluate the results. The Confusion matrix is shown in Figure 12.

$$F_{1}Score=2\frac{Precision\times Recall}{Precision+Recall}$$

(2)

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(3)

$$Precision=\frac{TP}{TP+FP}$$

(4)

$$Recall=\frac{TP}{TP+FN}$$

(5)

TP indicates a True Positive, where the prediction matched with what happened and the event was true.

TN indicates a True Negative, where the prediction matched what happened and the event was not true.

FP indicates a False Positive, where the prediction did not match what happened and the event was true.

FN indicates a False Negative, where the prediction did not match what happened and the event was not true.

4. Discussion

The mushroom dataset containing 2000 images was divided into two sets: a poisonous mushroom set containing 527 images and an edible mushroom set containing 1473 images. The dataset was divided into training and test sets in a ratio of 90:10.

4.1. Mushroom Classification Using the CNN

As is shown in

Table 4. CNN parameters used in training.

Parameters	Architecture
	AlexNet	ResNet-50	GoogLeNet	Proposed Model
Learning rate	0.001	0.001	0.001	0.001
MaxEpochs	10	10	10	10
MiniBatchSize	40	40	40	40

, in training, each architecture used a learning rate = 0.001, MaxEpochs = 10, MiniBatchSize = 40, and was processed with a single GPU.

Figure 13 shows the training results and validation test accuracy for each model. As shown in Figure 13a, AlexNet achieved a stable training accuracy after 160 iterations with an average training accuracy of 99.60%. As shown in Figure 13b, GoogLeNet achieved a stable training accuracy after 125 iterations with an average training accuracy of 99.70%. As can be seen in Figure 13c, ResNet-50 achieved a stable training accuracy after 40 iterations with an average training accuracy of 99.55%. In addition, Figure 13d shows that the proposed model achieved stable training accuracy after 340 iterations with an average training accuracy of 97.95%.

As can be observed in

Table 5. Analysis of the CNN test results.

Architecture	Accuracy	Time
ResNet-50	99.50%	5.50 min
GoogLeNet	99.50%	2.20 min
AlexNet	99.00%	1.42 min
Proposed Model	98.50%	1.10 min

, ResNet-50 was found to be the most accurate model with an accuracy of 99.50%, a precision of 99.33%, a recall of 100%, an F1 score of 99.66%, and a training time of 5 min 50 s. This was followed by GoogLeNet with an accuracy of 99.50%, a precision of 99.35%, a recall of 100%, an F1 score of 99.97%, and a training time of 2 min 20 s. Next came AlexNet with an accuracy of 99.00%, a precision of 98.61%, a recall of 100%, an F1 score of 99.30%, and a training time of 1 min 42 s. Finally, the proposed model exhibited an accuracy of 98.50%, a precision of 99.39%, a recall of 98.79%, an F1 score of 99.09%, and a training time of 1 min 10 s. The confusion matrix accuracy is shown in Figure 14.

4.2. Mushroom Classification Using the R-CNN

As is shown in

Table 6. R-CNN parameters used in training.

Parameters	Architecture
	AlexNet	ResNet-50	GoogLeNet	Proposed Model
Learning rate	0.001	0.001	0.001	0.001
MaxEpochs	1	1	1	1
MiniBatchSize	20	20	20	20

, in training, each architecture utilized a learning rate of 0.001, MaxEpochs = 1, MiniBatchSize = 20, and was processed with a single GPU. An example of detection using the R-CNN is shown in Figure 15.

As is shown in

Table 7. Analysis of the R-CNN test results.

Architecture	Accuracy	Time
ResNet-50	96.50%	13.44 min
GoogLeNet	96.50%	7.28 min
Proposed Model	95.50%	4.37 min
AlexNet	95.00%	5.07 min

, ResNet-50 was found to be the most accurate model with an accuracy of 96.50%, precision of 100%, recall of 95.24%, an F1 score of 97.56%, and a training time of 13 min 44 s. GoogLeNet followed with an accuracy of 96.50%; 99.30% precision; 95.92% recall; an F1 score of 97.58%; and a training time of 7 min 28 s. Next, the proposed model had an accuracy of 95.50%; 94.59% precision; 99.29% recall; an F1 score of 96.89%; and a training time of 4 min 37 s. Finally, AlexNet demonstrated an accuracy of 95.00%; 97.24% precision; 95.92% recall; an F1 score of 96.58%; and a training time of 5 min 7 s. The confusion matrix accuracy is shown in Figure 16.

5. Conclusions

In this study, a method for the classification of edible and poisonous mushrooms was proposed. The experimental results show that the proposed model can accurately classify poisonous and edible mushrooms and can shorten training and testing times. Moreover, the testing time and accuracy of the proposed model were compared with three pretrained models: AlexNet, ResNet-50, and GoogLeNet.

In the mushroom classification experiment carried out using CNN, the proposed model provided the result the fastest, in 1 min 10 s, while maintaining an accuracy level of 98.50%. This was less accurate than ResNet-50, and GoogLeNet at 99.50%; however, they took 5 min 50 s and 2 min 20 s, respectively.

In the mushroom detection experiment using R-CNN, ResNet-50 also demonstrated the highest accuracy level of 96.50%, but took the longest time for training, i.e., 13 min 44 s. The proposed model again produced the training results the fastest in only 4 min 37 s, with an accuracy of 95.50%. However, there was a different of 1% in the accuracy levels of ResNet-50 and the proposed model, and ResNet-50 took 9 min 7 s to obtain a result.

In the future, more varieties of mushrooms and different backgrounds can be added to the dataset to provide greater coverage and develop improved applications for mushroom classification.