1. Introduction
The technique of images segmentation involves breaking an image up into a number of distinct, non-overlapping parts and extracting the desired, human-interest-centered regions. From image processing to image analysis, it is a committed step that makes it easier for following computer vision, etc. There are other widely used image segmentation methods in addition to the Otsu threshold image segmentation method [
1,
2,
3] in the field. The maximum entropy method, for instance [
4]. A well-known edge detection strategy for extracting edge information from images is the Prewitt image segmentation technique [
5]. This method identifies the edges in the image by detecting gradient changes and is based on gradient calculation of the grayscale values of image pixels. The watershed image segmentation method [
6] is a region-based image segmentation method. In this method, the image is considered as a terrain map, where the height represents the grayscale value of the image. Image segmentation is achieved by finding watersheds, which divide the image into different regions or sets of different regions. The advantage of this method is that it can preserve areas with clear edges and rich textures in the image. However, there is also a drawback of over segmentation for images with large uniform areas. The Canny image segmentation method [
7] is a classic edge detection algorithm that can accurately detect edges in an image. This algorithm determines image edges by detecting gradient changes, and improves the accuracy and continuity of edges through non maximum suppression, dual threshold processing, and edge connectivity. The Sobel image segmentation method [
8] is a commonly used edge detection algorithm that is used to detect edges in images. The Sobel operator is used to calculate the gradient value of image pixels to determine the position and direction of the image. The Robert image segmentation method [
9] is a classic edge detection algorithm that applies the Robert operator to convolution the image and calculates the gradient values between adjacent pixels to determine the position and direction of the edges. An image segmentation technique based on the Laplacian Gaussian filter is the log image segmentation approach [
10]. By applying Gaussian and Laplacian filtering procedures to the image, this technique finds edges and other features in the image. The threshold image segmentation approach is a common one in image processing technology because of its benefits of simplicity, effectiveness, and strong robustness. Single-threshold image segmentation and multi-threshold image segmentation are two common threshold segmentation techniques. The threshold is determined using the Otsu threshold segmentation technique. The time complexity grows exponentially as the number of thresholds increases, as does the quantity of calculation.
One might think of the threshold image segmentation approach as an optimization issue. Researchers are increasingly merging threshold image segmentation with intelligent optimization algorithms as swarm intelligence optimization techniques grow over time. For instance, based on the enhanced Firefly algorithm [
11], better particle swarm optimization algorithm [
12], improved genetic algorithm [
13], improved whale optimization algorithm [
14], and improved cuckoo algorithm [
15]. You can use these clever optimization methods to identify appropriate optimization objective functions. The Otsu method and the maximum entropy method are common objective functions used in image segmentation.
Differential evolution [
16] (DE) was proposed by Storn and Price in 1997. Because of its small control parameters and strong robustness, differential evolution has been widely concerned by researchers. The basic operations of differential evolution include initial population, mutation, crossover and selection. As long as the population iteration conditions are met, the population will iterate continuously to find the optimal solution. Therefore, differential evolution is also a parallel search algorithm. In recent years, with the in-depth innovation of research, more and more researchers found that control parameters and mutation strategies are the main factors that affect the performance of differential evolution. The key control parameters in differential evolution mainly include population size NP, mutation operator F and crossover factor CR. Based on this, researchers continue to study and find variants of differential evolution. In reference [
17], AC Sanderson et al. proposed an improved differential evolution (JADE) using a new mutation strategy of selecting mutation vectors from an external archived population, and the control parameters F and CR in the population obey Cauchy distribution and Normal distribution, respectively. Ryoji Tanabe et al. [
18] developed an improved differential evolution algorithm (SHADE) based on JADE, which guides the selection of control parameters in subsequent iterations through the number of successful individuals with historical mutations. The experimental results showed that SHADE outperforms previous state-of-the-art DE algorithms on a large number of benchmark problems. Wu Deng et al. [
19] proposed a differential evolution with a new hybrid mutation factor and adaptive control parameters. Experimental results show that compared with other DE variants, the algorithm has better performance on the above test functions.
The differential evolution algorithm (DE) is a population intelligent optimization algorithm that searches for the optimal solution by simulating the process of natural evolution. With the continuous deepening of research on differential evolution algorithms, their application fields are gradually expanding. More and more optimization problems are combined with differential evolution algorithms to solve the optimal solution, such as scheduling problems [
20], traveling salesman problems [
21], and image processing [
22]. In scheduling problems, differential evolution algorithms optimize scheduling strategies and resource allocation to maximize work efficiency or minimize work time. Differential evolution combined with scheduling problems can handle complex constraint problems and quickly converge to find the optimal solution. The traveling salesman problem is a classic combinatorial optimization problem that requires finding the shortest path so that the traveling salesman can visit several cities and ultimately return to the starting city. The differential evolution algorithm searches for the optimal solution by adjusting the arrangement order of cities, with the goal of minimizing travel and finding the optimal solution for the shortest path. In this problem, differential evolution algorithm can efficiently search for the global optimal solution. In image processing problems, differential evolution algorithms are used for image registration, image segmentation, feature extraction, and image enhancement.
Bird swarm behavior is simulated via particle swarm optimization [
23]. In the process of segmenting a picture, each particle represents a potential segmentation scheme, and the best segmentation outcome is sought by continual iteration. With the help of pheromone propagation and reinforcement techniques, the Ant Colony algorithm [
24] eventually achieves the best segmentation outcomes by simulating the behavior of ants as they look for food. The genetic algorithm [
25] mimics the evolution process in nature by repeatedly iterating the population through selection, crossover, and mutation operations, maintaining its diversity and exploratory capacity through mutation, and finally discovering the best answer. Due to its few control parameters, robustness, and difficulty finding optimal solutions, the differential evolution method [
26] performs well and offers advantages in picture segmentation.
The differential evolution variant tested against pertinent benchmark functions performed really well. Researchers are increasingly using differential evolution to enhance and optimize image processing outcomes. To determine the ideal threshold and enhance the performance of image segmentation, Sushi L. Kumar et al. integrated differential evolution with the Otsu threshold segmentation method in reference [
26]. The experimental results demonstrate that the application of differential evolution to image segmentation can significantly increase the segmentation quality when compared to the segmentation results obtained using the conventional Otsu threshold segmentation method. However, the overall performance of differential evolution depends on the ability to balance global search and local search. The reason for its greater impact is the selection of control parameters. However, the above differential evolution proposed by Sushi L Kumar et al. is applied to threshold image segmentation, with high time complexity and poor robustness. Unable to meet the processing objectives of multi-threshold image segmentation. In reference [
27], Helen Vicente Humann Ayala et al. proposed an improved differential evolution (BDE) based on the generation of beta distribution adaptive control parameters F and CR values. In each iteration, the adaptive control parameters generate random F and CR values, enhancing the random search ability of the algorithm and improving the performance of the algorithm. Experimental results show that the improved differential evolution (BDE) is more effective than the FODPSO algorithm. However, due to the fact that the BDE algorithm randomly generates the values of control parameters F and CR, it cannot be ensured that the F and CR values are suitable for the entire iterative process of the algorithm. Based on the above algorithm. In this paper, a differential evolution based on adaptive double mutation is proposed for image segmentation. The population is divided according to the fitness value, and the double-mutation strategy is used to maximally balance the global search ability and local search ability of the algorithm. In addition, the values of control parameters F and CR are adaptively adjusted according to different stages of population evolution. The improved differential evolution in this paper is applied to Otsu multi-threshold image segmentation, and the maximum inter class variance method is regarded as the objective function to find the optimal threshold. The experimental results show that the algorithm is more accurate than other benchmark algorithms in image segmentation, with low time complexity and strong robustness.
Here, is the rest of the essay: The related work, such as the traditional differential evolution and Thresholding picture segmentation, are introduced in the second section. The third section introduces the TRDE algorithm proposed in this article. The fourth section introduces the experimental results, the fifth section is the analysis and research discussion of the experimental results, the sixth section is the references and future work, and the third part describes the improved double-mutation adaptive differential evolution algorithm (TRDE).
2. Related Work
2.1. The Classical DE Algorithm
Differential evolution is a population intelligent optimization algorithm. DE algorithm includes the following four steps: population initialization, mutation operation, crossover operation and selection operation. The following is the basic process of the DE algorithm:
- (1)
Population initialization
Randomly generate the initial population, initialize and generate NP individual vectors in d-dimensional space, denoted as:
Among them, NP is the population size, and g is the current evolutionary algebra. The NP individual vectors generated in d-dimensional space are randomly generated, and the specific formula is as follows:
In the above formula, represents the ith randomly generated individual, represents the Upper and lower bounds of the ith individual vector, represents a random number generated within the range of (0,1).
- (2)
Mutation
The DE algorithm achieves individual vector mutation through differential strategy, which is also the core idea of the DE algorithm. Randomly select several individual vectors from the population for differential operation to generate differential vectors. Common mutation strategies are as follows:
DE/best/2:
wherein
represent five randomly selected individual vectors from the population that are different from each other,
represents the variation vector of the ith individual in the generated generation, and F is the scaling factor.
- (3)
Crossover
Crossover operation is to cross the generated mutated individual vector with the parent individual vector to generate an experimental individual vector. The commonly used binomial cross operation in DE algorithm is as follows:
In the above formula, CR is the crossover operator, is a random integer selected from {1, 2, …, d} to ensure that the mutated individual has at least one individual component inherited to the next generation, avoiding the same vector as the parent individual.
- (4)
Selection
The selection operation in the DE algorithm adopts a greedy strategy, which selects the individual vector with the best fitness value as the evolutionary offspring vector through comparison. The specific formula is as follows:
2.2. The Otsu Threshold Segmentation Method
The maximum inter class variance method [
28] was proposed by Japanese scholar Otsu in 1997. The basic principle of Otsu for image threshold segmentation is to divide the image into background and target categories based on the grayscale characteristics of the image, calculate the number of pixels in the image, and solve for the maximum variance between the image target and background to achieve image threshold segmentation.
For a pair digital image of M × N, where i is the grayscale value, L is the grayscale level, and the range of grayscale values is [0, L − 1].
represents the number of pixels with grayscale level i, and the probability of grayscale level i appearing in the image is
, then:
Based on the grayscale characteristics of digital images, the pixel points in the image are divided into two categories: target and background using threshold s, Represented by
and
. The grayscale values between [0, s] are classified as
, and the remaining grayscale values between [s + 1, L − 1] are classified as
. For the entire image, the average grayscale value is:
The mean of
and
is:
Based on the above formula, we can obtain:
The definition of inter class variance is as follows:
Through the above formula, we can see that the s value corresponding to the maximum inter class variance is the optimal segmentation threshold.
2.3. Existing Problems in Research and Motivation for Improvement
The segmentation criterion for traditional Otsu is the largest variance between the target and background of the picture, which divides the target and backdrop of the image into two categories. However, the histogram of the segmented Grayscale typically does not exhibit a bimodal form, making it challenging to obtain precise image segmentation findings. To obtain the ideal threshold, many researchers treat threshold picture segmentation as an optimization problem and pair it with clever optimization methods. The control settings and mutation approach in the traditional DE algorithm have the most effects on the algorithm’s overall performance. In the DE algorithm, the important control parameters are the scaling factor F and the crossover factor CR. The scaling factor F controls the magnitude of differential evolution vector, which affects the convergence ability and Rate of convergence of the algorithm. The crossover factor CR affects the probability of generating a vector of experimental individuals in crossover operations to inherit genes from their parents or mutated individuals. The second part is the improvement of mutation strategy, balancing the global search ability and local search ability of the algorithm. To sum up, this paper proposes an adaptive double-mutation differential evolution algorithm (TRDE), which divides the population according to the fitness value of the population, adopts corresponding mutation strategies for different sub-populations, and adaptively adjusts the values of control parameters F and CR to adapt to different stages of population evolution.
3. The TRDE Algorithm
Researchers have enhanced and created variants of the DE algorithm based on it in order to significantly balance the global and local search capabilities. There are primarily two components in the DE algorithm that greatly influence how well it performs. The first step is to adjust the values of the parameters F and CR. The scaling factor F determines the size of the differential evolution vector and has an impact on the algorithm’s capacity for convergence and rate of convergence. The possibility of creating a vector of experimental individuals in crossover operations who would inherit genes from their parents or modified individuals is influenced by the crossover factor CR. a variety of adaptive parameter control DE algorithm versions, such as in reference [
29]. A differential evolution algorithm based on dual mutation strategy is proposed for image segmentation. The experimental results show that this algorithm is competitive with other improved DE algorithms in terms of optimization accuracy and convergence speed. Vladimir Stanovov et al. proposed a method in reference [
30] to adjust the values of control parameters F and CR based on fitness values, and increase the biased Lemmer mean and LBR to improve its performance. The experimental results show that the algorithm performance has been greatly improved. The second part is the improvement of mutation strategy, balancing the global search ability and local search ability of the algorithm. Regarding the improvement of mutation strategy, there are also many variants of DE algorithm. For example, in reference [
31], Wan-li Xiang et al. proposed an enhanced DE algorithm (EDE), which uses a reverse learning strategy to increase the diversity of the initial population, improves the mutation strategy on the basis of JADE algorithm, and improves the overall performance of the algorithm. As mentioned in reference [
32], Ali Wagdy Mohamed proposed a mutation strategy based on triangular mutation rules, which greatly balances the global detection ability and local search ability of the population.
3.1. Dual Mutation Strategy
The main component of the DE algorithm is the mutation strategy, which can make the population’s individual evolution vectors and orientations display diversity. Based on the fitness scores of various population vectors, this article separates the population into two subpopulations. Mutation strategy 1 is used for populations with high fitness values, while mutation strategy 2 is used for populations with low fitness values. In general, it is simpler to locate the ideal solution close to people with high fitness values or those with low principle fitness values. Consequently, the dual mutation technique can balance population diversity and local search capability, speeding convergence.
After initializing the population, calculate the population fitness value, and divide the population into two subpopulations based on the fitness value. Subpopulation 1 stores the vector of individuals with high fitness values, while Subpopulation 2 stores the vector of individuals with poor fitness values. Improved mutation strategy 1:
is a random selection of one of the top d% individual vectors from Subpopulation1 as the basis vector. In this article, the selection of the d parameter adopts a linearly decreasing function. The specific changes in parameter d are as follows:
This article set
, gen is the current population evolution algebra, and generations is the total evolution algebra. The variation of d parameter shows a decreasing trend with the increase of iteration times. As the number of iterations in the population continues to increase, the fitness values of individual vectors in the population continuously concentrate in a certain local region. Therefore, in order to improve the local convergence ability of the population and find the optimal solution, in the later stage of iteration, the selection of the optimal d% individual vector gradually decreases, accelerating population convergence and enhancing convergence ability. Generally speaking, individuals with high fitness values or poor principle fitness values are more likely to find the optimal solution, while individuals with poor fitness values can enhance population diversity. Therefore, the mutation strategy 2 proposed in this article is as follows:
Based on the fitness value, randomly select four distinct integers in Subpopulation2, and generate a difference vector for mutation operation to improve population diversity.
3.2. Parameter Adaptation
The fixed values of the control parameters F and CR in the traditional DE algorithm make it impossible for them to adapt to the algorithm’s needs at various phases of evolution. As a result, adaptive parameter control techniques are being used by more and more DE variations. In general, a DE algorithm should have strong global search capabilities, retain population variety as much as feasible in the early stages of population evolution, and look for the world’s best answer. In the later stage of population evolution, as the number of iterations keeps rising, it becomes increasingly important to have strong local convergence capacity and accelerate convergence.
The scaling factor F in the DE algorithm is closely related to the Rate of convergence, which controls the amplitude of the difference vector. In the early stages of population evolution, a larger value of the scaling factor F is beneficial for maintaining population diversity and ensuring that the algorithm searches throughout the entire solution space. As the number of iterations continues to increase, in the later stage of evolution, the optimal solution of the individual vector gradually concentrates in a certain local region. It is necessary to gradually reduce the value of F and strengthen the local convergence ability of the algorithm. The crossover factor CR affects the probability of experimental individuals generated by crossover operations inheriting genes from mutated individuals or parent individuals. Generally speaking, in the early stages of algorithm evolution, to ensure the algorithm’s global search ability, the CR value is small. In the later stage of algorithm evolution, in order to improve the local search ability of the algorithm with a larger CR value, combined with changes in the fitness value of individual vectors in the population, this paper proposes adaptive parameter control as follows:
where gen represents the evolutionary algebra of the current population, and generations represents the total number of iterations,
represents the maximum value for setting the scaling factor,
represents the minimum value for setting the scaling factor, This article set
.
The changes in the scaling factor F proposed above show a decreasing trend, which is consistent with the convergence characteristics of differential evolution.
where gen represents the evolutionary algebra of the current population, and generations represents the total number of iterations,
represents the maximum value for setting the crossover factor, and
represents the minimum value for setting the crossover factor. In this article, set
. The change in CR in the above formula shows a monotonic increasing trend, which greatly balances the contradiction between population diversity and Rate of convergence. From the test results, it can be seen that the improved adaptive control parameter method has advantages over other benchmark algorithms.
3.3. Pseudo Code of TRDE
1: | Initialize population p individuals and calculate their fitness values, NP = 30, gen = 1, generations = 100; |
2: | for gen = 1 to generations do |
3: | | for i = 1 to NP do |
4: | | |
|
5: | | |
|
6: | | | Implement mutation in Equations (18) and (20); |
7: | | | Implement crossover in Equation (7); |
8: | | Implement selection in Equation (8); |
9: | | end for |
10: | | gen = gen + 1; |
11: | | end for |