Research On Image Method Of Crop Disease Under Sparse Representation Framework

Crop disease is one of the major agricultural disasters in China.Crop disease recognition plays key role in controlling crop diseases.To solve the problems that the existing crop disease recognition methods mainly focus on a single crop and their recognition accuracy still cannot meet the requirement for application,this paper systematically study the crop leaf disease recognition method of wheat,corn,peanuts,and cotton,by using the knowledge of computer vision,image processing,and pattern recognition,etc.The main works are as follows:1)We collect 422 crop leaf disease sample images in North china,which contain 22 kinds of common diseases of wheat,corn,peanuts,and cotton.Based on the leaf and disease spot segmentation result of each disease image,the leaf and disease spot features,which respectively characterize the crop and disease type,are firstly extracted.Then,the two kinds of features are combined and normalized to obtain the final data feature vector of each crop disease image.As a result,all the data feature vectors of the crop leaf diseased images are used to construct a crop leaf disease data set.Moreover,a software for crop leaf disease processing and feature extraction is also developed based on the above-mentioned research results.2)Experiments show that the recognition rate using original feature of crop leaf disease is terrible.To improve the crop disease recognition rate,we adopt feature selection to remove redundant information and interference information in the original features.For such reason,we propose an 2,pL-norm based discriminant locality preserving Projections(DLPP_L2,p)algorithm for feature selection.As we know,In DLPP,2L norm is employed to measure the distance between the samples,which is likely to amplify the impact of the outliers.Additionally,the features obtained by DLPP is a linear combination of the original features,and the redundant information cannot be removed.As a result,there may still be some redundant information or interference information in the extracted features,and directly using the DLPP features may degrade the recognition performance.Compared with DLPP,DLPP_L2,p adopts 2,pL-norm to measure the distance between the samples so as to reduce the impact of the outliers.Furthermore,DLPP_L2,p also impose a row sparse constraint on the mapping matrix to achieve the purpose of feature selection.3)The collaborative representation-based classifier has been widely used in the field of image recognition.However,when there are outliers in training set,its performance will be greatly degraded.In addition,researches show that effects of data features and local relationships of samples are critical to data classification.Therefore,we propose a double-weight collaborative representation-based classifier for crop leaf disease recognition.Experimental results on our crop leaf disease dataset demonstrate the superiority of our method compared with some popular classifiers,such as SVM,Neural Network,Sparse Representation classification-based classifier,Collaborative Representation classification-based classifier,and Regularized Robust Coding-based classifier,etc.