节点文献

基于模糊集合理论的颗粒目标分割和识别

Segmentation and Recognition of Particle Object Based on Fuzzy Set Theory

【作者】 尹诗白

【导师】 赵祥模;

【作者基本信息】 长安大学 , 交通信息工程及控制, 2013, 博士

【摘要】 基于图像处理的颗粒目标自动分割和识别是工农业生产中实现自动化检测的重要方法,利用模糊集合理论可以有效地解决目标分割和识别中的模糊问题,提高分割识别精度。本文针对模糊集合理论在颗粒目标分割和识别中的关键问题进行了深入研究,主要工作如下:1.针对模糊划分熵多阈值分割算法在确定最优阈值时易陷入局部最优解的问题进行分析,认为这是由于常用的遗传算法对模糊划分熵函数寻优时,易早熟的现象导致的。针对该问题,采用染色体编码、遗传参数设置、进化方向等方面改进后的遗传算法进行寻优,应用了一种改进遗传算法的模糊划分熵多阈值分割算法对颗粒目标图像进行分割。实验结果验证了该算法在颗粒目标图像分割上的可行性和优越性。2.针对模糊划分熵多阈值分割算法对目标像素比例大于10%的颗粒目标图像分割时,存在的噪声大、空间相关性差、运行时间随阈值数量增加而线性增长的问题,提出了一种递推的模糊划分熵多阈值颗粒分割算法。首先将图像总模糊熵公式中不同变量的组合计算转化为递推过程;然后保存部分不重复的递推结果用于后续计算;最后使用种群寻优算法快速确定最优阈值,并采用基于区域的图割算法对分割结果实施优化。实验部分首先对不同的种群寻优算法进行性能评估;然后将递推算法与不同的种群寻优算法相结搜索全局最优阈值,以验证递推算法的有效性;最后将此算法与常用的多阈值分割算法相对比,以证明提出算法的优越性。通过多幅骨料颗粒目标图像分割测试表明,运行时间随阈值数量的增加基本保持稳定,相比于无递推策略的模糊划分熵多阈值分割算法,运行时间提高约10倍-100倍,且较好地去除了噪声,强化了空间相关性,提高了分割精度。3.针对模糊划分熵多阈值分割算法对目标像素比例小于10%的模糊小颗粒目标图像分割时,存在的噪声大,背景干扰强,分割精度低的问题,提出了一种自适应模糊划分熵多阈值颗粒分割算法。首先采用迭代验证法确定隶属度函数窗宽;然后使用自动的图像划分算法将图像分为若干子图;最后采用基于人工蜂群寻优的递推模糊划分熵多阈值分割算法对各子图进行分割,并使用基于像素的迭代图割算法对分割后的结果实施优化。通过多幅人工仿真图像和真实FISH基因图像分割测试表明,常用的自适应分割算法和其它寻优策略的模糊划分熵多阈值分割算法的错误划分概率均大于8.00×10-2,而本文算法的错误划分概率小于7.00×10-2。4.针对不同颗粒目标的特征量具有模糊界限的问题,提出了基于模糊综合评判的目标识别方法。首先对分割优化后的结果进行形态学图像预处理;然后选择适当的特征量,建立相应的隶属度函数;最后确定各特征量的权重值,建立模糊关系矩阵,进行模糊综合评判。通过多幅骨料颗粒和FISH基因颗粒目标图像测试表明,本文算法的识别正确率大于95%。5.构建了一个较为完整的颗粒目标自动分割和识别系统。在系统的构建过程中,为了对不同种类的颗粒目标图像实现模糊划分熵分割算法的自动选择,采用模糊神经网络的方法来解决,即将颗粒目标图像在不同模糊划分数下获得的模糊熵值作为输入,各因素的模糊综合评判矩阵作为神经元,确定的模糊划分熵方法作为输出,经过大量样本的反复学习,最终获得基于模糊神经网络的自动分割方法选择模型。通过多幅颗粒目标图像的运行实例,验证了系统的有效性。

【Abstract】 The automatic particle segmentation and recognition is widely used in industry andagriculture. It is an important method of automatic detection. In this work, fuzzy set theory isapplied and well solves the fuzzy problem of particle segmentation and recognition. It wellhelps to improve the accuracy of segmentation and recognition. Several problems of fuzzy settheory and its application in particle segmentation and recognition are studied intensely. Themain works are arranged as follow:1. The fuzzy partition entropy has been widely adopted as a global optimizationtechnique for finding the optimized thresholds for multilevel image segmentation. However, itusually involves the local optimum thresholds. The main reason is the applied geneticalgorithm easily lead to the premature phenomena. In order to solve this problem, theimproved GA has been adapted which improves encoding mechanism, genetic operators andevolutionary direction of conventional genetic algorithm. Many experiments are carried outon particle images to demonstrate the feasibility and efficiency of the improved scheme.2. The fuzzy partition entropy segmentation involves expensive computation as thenumber of thresholds increases and often yields noisy segmentation results since spatialcoherence is not enforced. In this paper, an iterative calculation scheme is presented forreducing redundant computations in entropy evaluation. The efficiency of threshold selectionis further improved through utilizing population optimization algorithm. Consequently,instead of performing threshold segmentation for each pixel independently, the presentedalgorithm over-segments the input image into small regions and uses the probabilities offuzzy events to define the costs of different label assignments for each region. The finalsegmentation results are computed using graph cut, which produces smooth segmentationresults. Experiment results indicate that the presented method is not only superior to the samefuzzy entropy methods with different optimizing strategies in terms of processing time, butalso outperforms widely-used multi-threshold segmentation methods in terms of thesegmentation quality of aggregate images. Furthermore, the iterative scheme can dramaticallyreduce the runtime and keep it stable as the number of required thresholds increases. Dependson the optimization methods and the number of thresholds, the speedup varies from10to100times.3. To solve the problem of noise and poor precision about fuzzy partition entropy approach for segmenting the small particle objects (the propotion of object pix number ismore than10%), a new adaptive fuzzy partition entropy algorithm for the multi-thresholdsegmentation is proposed. Firstly, a threshold zone is obtained automatically by the iterativevalidation algorithm. After that, the particle images are divided into several differentsub-images by the automatic block algorithm. Finally, the fuzzy partition entropy based onthe population optimized algorithm is adopted to find out the best thresholds for eachsub-image. And the segmentation results are computed using graph cut for smoothing thesegmentation results. After being evaluated by various types of real FISH images andsimulated images, the misclassification error of other common algorithms are above8.00×10-2,while the one of proposed algorithm is less than7.00×10-2.4. To solve the fuzzy boundary problem of different characteristic parameters, a fuzzycomprehensive evaluation algorithm for particle recognition is presented. Firstly, themorphological image preprocessing method has been adopted to eliminate the adhesionparticles. After that, the characteristic parameters are extracted from the candidate particleobjects, and corresponding membership functions are built up. Finally, the fuzzycomprehensive evaluation has been applied to identify the particle objects in combinationwith characteristic weights and fuzzy relation matrix. The experiment results show that theproposed method produces more accurate results than the results acquired by the commonrecognition methods and the accuracy is up to95%.5. A completed particle segmentation and recognition system is designed. The fuzzyneural network is adopted to determine the suited fuzzy partition entropy method for differentparticle images. In order to implement the function of automatic decision, the fuzzy entropyvalues which under different fuzzy partition numbers are set as the input of the neural network.The fuzzy comprehensive evaluation matrix is used as nerve cell. And the determined fuzzypartition entropy method is designed as output. After learning a large number of samples, themodel of automatic algorithm selection can be acquired. Experiment results show that thesystem can work effectively in particle segmentation and recognition.

  • 【网络出版投稿人】 长安大学
  • 【网络出版年期】2014年 07期
节点文献中: