Segmentation Method Of Crop Disease Images Under Complex Background Based On Improved LBP

Segmentation of crop disease leaf is a hard problem in computer vision and image analysis based identification of crop diseases,is to extract the significant interest lesion from the original disease leaf image,exclude non-significantly less-important area and prominent the lesion image.It is in favor of the later detection,diagnosis and identification of crop disease.Disease leaf image segmentation is an important topic in image processing task and is a key step in the crop leaf image based disease and pests recognition methods.It is becoming more and more important in the agricultural automation,such as in the field of crop disease automatic recognition.Although there are many methods of crop disease leaf image segmentation,many methods cannot meet the practical needs of current crop disease identification system.Threshold is one of the representative widely used methods,which depends on the feature map and threshold value.In the threshold segmentation method of crop disease leaf image,there is an assumption that the original image can be segmented.That is to say,clusters in the color histogram correspond to spot region that can be extracted by separating these histogram clusters.However,the fact disease leaf image is very complex containing a lot of color components,so there are no obvious valley points in the histogram of the disease leaf image,which results that many traditional image segmentation methods cannot be applied directly to the diseased leaf image segmentation.The image preprocessing process such as filtering,smoothing and enhancement can improve the segmentation effect.With its ability of discrimination and simplicity,Local Binary Pattern(LBP)has become a popular approach in many fields,especially in computer vision,and it is easy to implement in real-time systems.Although LBP has been applied effectively to image segmentation,it is less effective in complex disease leaf segmentation.As for the problem of leaf image segmentation for crop disease recognition,this paper studies LBP and its improved algorithms for the segmentation of the actual disease leaf images.In the plant disease recognition,if we obtain the accurate and complete segmentation of lesion image,we can accurately represent and descript the diseases image for disease recognition.Because of the complex diversity of the disease blade image,many image segmentation methods are rarely satisfied.To solve this problem,we propose an improved LBP algorithm,which is called modified adaptive center-symmetric LBP(MACLBP),and propose a segmentation method for disease leaf image by combining with the Otsu threshold segmentation algorithm.The main contributions are as follows:(1)Redefine the traditional LBP pattern classification method according to the color and texture features of the disease blade image.(2)Define a block color correlation metric S,if a block of S value is small,the block is directly classified as non-lesion without segmenting,and the blocks are classified into meaningless local mode.This metric can characterize the distribution information of color spot disease in leaf color space,but also make up for the neglect of different colors of lesion in the color histogram correlation;(3)An improved LBP algorithm(MACLBP)is proposed.The algorithm uses the correlation between each pixel and the adjacent pixels,using adaptive center symmetric LBP(ACS-LBP)algorithm to get the texture feature histogram and the weights of different regions,the weighted histogram texture feature vector of lesion is obtained through the weighted connections;(4)The LBP feature image is segmented using Otsu algorithm,so as to get the two-value lesion image.The algorithm is according to the gray level distribution of the image characteristics,and the disease image is segmented by a threshold selected criteria of maximizing the variance between the spot part and normal leaf part.Through the above improvements,the proposed disease leaf image segmentation method has a fast calculation speed,strong anti-noise,robustness to illumination conditions,and without prior segmentation of disease leaf area,and the lesion image can be effectively segmented.The proposed method considers not only the difference between the center pixel and neighborhood pixels into account,but also considers the relationship between the pixel and its neighborhood points,and can give prominence to the image intensity and pixel gray value changes,which can increase the contrast between the health and lesion,and in crop disease leaf images and normal green the boundary the contour information is more obvious.The experimental results show that the proposed algorithm can accurately and completely segment lesion,and does not require pre-segmentation of the disease leaf image,and provides an effective means for the late quantitative description of disease leaf image and disease recognition application.