Detection Of Cucumber Powdery Mildew Based On Support Vector Machines

The study on automatic cucumber powdery mildew detection using image processing is of important significance for recognizing cucumber powdery mildew accurately, reducing the disease on cucumber and ensuring national food quality. However, as a newly emerged field of application, the technique of cucumber powdery mildew detection based on image processing still has not been systematically studied yet. After reviewing some related research works on crop disease detection, an approach is based on image segment of cucumber leaf images, then color feature, shape feature and texture feature extraction, and training a support vector machine (SVM) based leaf disease detection. The system for cucumber powdery mildew detection using MATLAB was also designed and developed, which provided the basic platform for the study and development of the intelligent system of yield cucumber disease. The contents of the study could be briefly summarized as follows:(1) The segmentation algorithm of cucumber powdery mildew image was studied. To deal with the dimension difference among cucumber powdery mildew spots, multiple scales of watershed segmentation method were used. In addition, the watershed segmentation algorithm is based on the likelihood ratio of disease spot color and normal leaf color instead of the image intensity.(2) Color features, shape features and texture features for cucumber powdery mildew detection was studied. After the comprehensive consideration of various existing crop diseases identification approaches, we define some proper features suitable for cucumber powdery mildew spot detection. We proposed to not only consider the disease spot sample area, but also consider the surrounding pixels of the sample area to avoid the problem of inconsistent color caused by different sample collection conditions and different types of cucumber leaves. In addition, the local binary pattern (LBP) descriptor is selected to represent the texture feature cucumber powdery mildew spots, which was seldom used in existing works.(3) The approach for cucumber powdery mildew detection was studied. And a detection method based on powdery mildew spots detection and powdery mildew leaf detection using the output of powdery mildew spot detector was proposed. To reduce the training time, improve the detection accuracy, and make the classification model have better generalization ability, based on the comprehensive consideration of samples and the performance of classifier, we proposed the iterative training method to train the support vector machine based classifier, and train two kinds of cucumber powdery mildew spot detector based on radial basis function and the linear function respectively, and use grid search and cross validation to optimization the SVM parameter selection. In order to avoid over-fitting, we chose a linear SVM to detect powdery mildew leaves based on the output of powdery mildew spots detection. Finally, we choose the leaf detection based on the output of radial basis function.(4) Design the cucumber powdery mildew detection system.