显著性检测方法及其在黄瓜病害图像分割中的应用研究

发布时间：2019-07-03 10:14

【摘要】：最近几年,图像显著性检测是计算机视觉领域研究的热点。图像显著性检测的目的是能够将图像中感兴趣的目标区域自动地检测出来。对目标区域的检测精度与检测效率将直接影响到后续目标识别的性能。本文围绕如何提高显著性检测算法的精度和检测效率展开相关的理论方法研究,并将提出的显著性检测算法在黄瓜病害图像处理中进行了应用研究。论文的主要研究工作如下:(1)提出了一种基于先验信息和双权重的显著性检测算法(Saliency detection algorithm based on prior information and double weights,P I DWSD)。PIDWSD 算法主要是为了解决上下文感知显著性检测算法(Context-Aware saliency detection,CA)中存在的边缘丢失及检测精度不高的问题。PIDWSD算法首先使用超像素将图像分块,以获得良好的目标边缘;其次,引入高斯权重和欧氏距离权重,以获取精细化的显著图;接着,引入中心先验和非显著关联先验,以去除背景中的干扰信息;最后,通过非线性作用函数Sigmoid对得到的显著图进行调整优化。在Berkeley和MSRA1000数据库上进行测试。与其它显著性检测算法相比,该方法不仅能很好地解决边缘丢失问题,检测精度达到93%,而且具有较低的算法时间复杂度。(2)提出了一种融合流形排序和能量方程的显著性检测算法(Saliency detection algorithm combining manifold ranking and energy equation,MREESD)。该算法主要是为了解决传统显著性检测算法检测精度不高且显著种子选取鲁棒性不足的问题。首先,使用超像素方法将图像分块,提出了一种新的超像素间权重计算方法和显著种子选取方法,以增强算法的鲁棒性;其次,通过流形排序计算,以获取较优的显著图;为使得显著图更加精确,利用能量方程对得到的显著图进行平滑调整;对调整后的显著图进行阈值分割,将得到的二值图像与原图像进行掩码运算,得到最终分割结果。在MSRA1000图像显著性检测数据库上进行测试,准确率-召回率曲线显示在相同召回率下准确率高于其它算法,并且具有较高的F-measure值。最后,将MREESD同PIDWSD进行了实验对比,从实验结果中看出,MREESD算法具有更强的鲁棒性。(3)作物病害图像分割精度对病害自动识别效果具有关键作用。针对复杂背景下黄瓜叶部病害分割精度不高的问题,本文将显著性检测应用于自然环境的黄瓜叶部病害的图像处理中。首先,通过显著性检测算法提取出黄瓜病害叶片;其次,利用超绿特征对病害叶片进行处理,以扩大绿色正常部分和非绿色病斑部分的灰度差距,通过阈值分割出病斑;最后,利用形态学膨胀操作对得到的病斑进行处理,以获取更加饱满的病斑。实验结果表明,本文所提的算法在提取出的病斑上更加精确,错分率均低于5%。通过对黄瓜典型的四种病害进行分析,提取病害特征;最后,采用BP神经网络分类器对黄瓜病害进行分类识别,识别率达到83%以上,从而验证了本文所提的显著性检测算法在病害图像处理中的可行性和实用性。
[Abstract]:In recent years, image saliency detection is a hot topic in the field of computer vision. The purpose of the image saliency detection is to be able to automatically detect the target area of interest in the image. The detection accuracy and the detection efficiency of the target area will directly affect the performance of the subsequent target recognition. This paper studies on how to improve the accuracy of the significance detection algorithm and the detection efficiency, and applies the proposed significance detection algorithm to the image processing of cucumber diseases. The main research work of the thesis is as follows: (1) a significance detection algorithm based on a priori information and a double-weight is proposed (Salience detection algorithm based on priority information and double weight, P I DWSD). The PIDWSD algorithm is mainly to solve the problem of low edge loss and low detection accuracy in the context-aware significance detection algorithm (CA). The PIDWSD algorithm first uses the super-pixel to block the image to obtain a good target edge; secondly, introducing the Gaussian weight and the Euclidean distance weight to obtain a refined saliency map; then, introducing a center prior and non-significant correlation a priori to remove the interference information in the background; and finally, And the obtained saliency map is adjusted and optimized by the non-linear action function Sigmoid. Testing was performed on the Berkeley and MRA1000 databases. Compared with other significance detection algorithms, the method not only can well solve the problem of edge loss, the detection accuracy reaches 93%, but also has lower algorithm time complexity. (2) A significance detection algorithm for the ordering and energy equation of a fusion manifold (MREESD) is proposed. The algorithm is mainly used to solve the problem that the traditional significance detection algorithm is not high in detection precision and is not sufficiently robust to select a significant seed. Firstly, the super-pixel method is used to block the image, a new method for calculating the weight between the super-pixels and a method for selecting a significant seed is proposed, so that the robustness of the algorithm is enhanced; secondly, the optimal saliency map is obtained by the manifold sorting calculation, so that the saliency map is more accurate, And performing a threshold segmentation on the adjusted saliency map, and performing mask operation on the obtained binary image and the original image to obtain a final segmentation result. On the MRA1000 image significance test database, the accuracy-recall rate curve shows that the accuracy rate is higher than other algorithms at the same recall rate, and has a higher F-mean value. Finally, the MREESD is compared with the PIDWSD, and it can be seen from the experimental results that the MREESD algorithm is more robust. (3) The image segmentation accuracy of crop disease plays a key role in the automatic recognition of disease. Aiming at the problem of low segmentation precision of the cucumber leaf part under the complex background, the method is used in the image processing of the disease of the cucumber leaf part of the natural environment. firstly, a cucumber disease blade is extracted by a saliency detection algorithm; secondly, the disease blade is treated by using the super-green characteristic to expand the gray difference of the green normal part and the non-green disease spot part, and the disease spot is divided by a threshold value; and finally, The acquired disease spot is treated by the morphological expansion operation so as to obtain a more plump disease spot. The experimental results show that the proposed algorithm is more accurate in the extracted lesions, and the error rate is less than 5%. By analyzing the four diseases typical of the cucumber, the disease characteristics are extracted; and finally, the BP neural network classifier is adopted to classify and identify the cucumber diseases, and the recognition rate is more than 83 percent, So that the feasibility and the practicability of the significance detection algorithm in the disease image processing are verified.
【学位授予单位】：南京农业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：S436.421;TP391.41

【参考文献】