基于演化算法的水果图像分割

为了满足水果采摘机器人对图像分割算法实时性和自适应性的要求,在传统演化算法的基础上,提出了一种基于蜂王交配结合精英选择、截断选择分阶段的改进演化算法对水果图像进行分割。在设计选择策略时,将迭代过程划分为前中后3个阶段,分别采用蜂王交配算法、精英选择策略和截断选择策略来进行适应值的选择,这样既保证了种群的多样性,又克服了传统演化算法局部最优、收敛过快的缺点。试验结果表明,该文提出的水果图像演化分割算法无论从稳定性、分割效果,还是全局最优收敛速度上,都明显优于传统演化算法,分割的阈值稳定在3个像素之内;与Otsu算法、贝叶斯分类算法、K均值聚类算法、模糊C均值算法等其他算法相比,水果图像演化分割算法分割效果最好,对同一幅图像进行分割得到的分割识别面积参考值最大,而且运行速度最快,平均运行时间为0.08735 s,远少于其余4种算法;并能用于柑橘、荔枝、苹果等各种水果的图像分割,具有一定的通用性,达到水果采摘机器人视觉实时识别的要求,为水果图像分割及其实时获取提供了一种新的基础算法。

Fruit image segmentation based on evolutionary algorithm

Abstract: In order to meet the demand of real-time image and adaptive processing algorithms for picking robots, an improved evolutionary algorithm based on queen mating combined with elite choices and truncated choices by stages was proposed for fruit image segmentation,. The 8 bit binary code was used to correspond with the gray value of the fruit image in the improved evolutionary algorithm. The number of the initial population was set to 12 in the phase of the population initialization and the corresponding individual values ranging between 0 and 255 were generated by the random function. The twelve random numbers were the initial values of the evolutionary algorithm. Then an improved Otsu algorithm formula was selected as the fitness function. In the selection phase the iterative process was divided into before, middle and after stage, which were respectively used by queen mating algorithm, elitist choices strategy and truncated choices strategy to select the fitness value. In the first stage, the individuals were produced by a random function and then the best individual (queen) of the evolutionary algorithm was hybridized with the rest individuals (including the randomly generated individuals) to generate new individuals, and finally the individuals of the smallest fitness value were replaced by the new individuals. In the second stage, the elitist choices strategy was used and the individual of the smallest fitness value in the current generation was replaced by the individual of the most fitness value in the previous generation. In the third stage, the truncated choices strategy was used and the last half of individuals of the smallest fitness value in the current generation were replaced by the same number of individuals of the most fitness value in the previous generation. In this way, it could not only ensure the diversity of population, but also overcome the disadvantage of local optimized and too fast convergence using the traditional evolutionary algorithm. In the crossover phase, a single-point crossover method was used. In the mutation phase the selected mutation probability was 0.2, which was obtained by comparing the results of different experiments. In the termination phase the termination condition of the evolutionary algorithm in this paper was that the number of the current iteration had reached the maximum number set by the user in advance. The experimental results showed that, the proposed fruit image evolutionary segmentation algorithm was superior to the traditional evolutionary algorithm, and was better in terms of stability, segmentation effect and running speed, etc, and the segmentation threshold value was stabilized within the 3 pixels. Compared with the Otsu segmentation algorithm, the K-means clustering segmentation algorithm, the fuzzy C-means clustering segmentation algorithm and Bayesian classification segmentation algorithm, the fruit image evolutionary segmentation algorithm had the best segmentation effect and took the least run time. The average run time of the evolutionary algorithm was 0.08735 seconds, which was less than the other 4 algorithms. The evolutionary segmentation algorithm could be used for image segmentation of citrus, litchi, apple and other fruits, and so the algorithm had certain universal utility. The algorithm meets the demand of vision real-time positioning when being used for the fruit picking robot and can provide a new basis algorithm for the image segmentation and its real-time research.