Research On Key Techniques Of Crop Leaf Diseases Recognition Based On Image Retrieval

Crop diseases can lead to the decline of crop quality and yield,which brings huge economic losses to the farmers.Therefore,it is of great practical significance to recognize and diagnose crop diseases timely.Crop disease recognition based on image retrieval has great practical value and bright application prospects due to combining the advantages of common crop disease recognition technology and image retrieval technology.However,there are still some issues that need to be further studied and solved.For example,for the lesion segmentation in the process of crop disease-recognition based on image retrieval,the existing researches mostly used the leaf diseased images with simple background,which were not suitable to process the diseased images collected in the actual farmland environment.In the aspect of feature extraction of diseased images,only traditional low-level features were employed to describe the attributes of the lesions in the image,which ignored the higher-level image information.For the recognition of diseased images,the existing methods based on image retrieval technology didn't achieve good performance of the disease recognition because of using only a single feature in retrieval which couldn't completely describe the attribute of the diseased image.In order to solve above problems,this thesis focuses on the method study of the lesion segmentation from crop leaf diseased images in the farmland production environment,and researches the extraction method of higher-level abstract characteristics of diseased images.Finally,this thesis proposes a fusion method of image retrieval to recognize and diagnose crop diseases.The main work of this thesis can conclude as follows:1.This thesis adopts a two-layer framework to segment the crop leaf diseased images achieved in the farmland production environment effectively.Firstly,a saliency detection method is introduced to remove the background of the diseased images,which can remove most of the background without losing any lesion.Then the methods of graph theory,thresholding,pixel classification and K-means clustering are studied and compared,which realize the efficient extraction of the lesions from the diseased images without complex background.2.This thesis extracts higher-level abstract characteristics of diseased images from both global features and local features.Based on the traditional low-level visual features,a bag-of-lesions model is researched to represent the diseased image as a global histogram vector.And the common local feature,Sift,is introduced and combined with the bag-of-words model to represent the local attributes of the diseased images.These two characteristics are used in the follow-up disease recognition,which can provide a complete difference information between disease categories.3.This thesis studies a novel method for the disease identification based on image retrieval fusion.Firstly,the BoW_Traditional characteristic and the BoW Sift characteristic are extracted respectively to describe the global and local attributes of the diseased images completely.Then according to the similarity ranking rules of different features in image retrieval,an adaptive fusion strategy is proposed to integrate the retrieval results of these two characteristics effectively.This method achieves an average recognition accuracy of 90.84%,which has certain application value for the recognition and diagnosis of crop diseases in the actual environment.4.On the basis of the above research,this thesis develops a crop diseases recognition system based on image retrieval.Through the above research,this thesis realizes the accurate recognition of crop leaf diseases under the actual environment,which can provide technical support for the scientific diagnosis and prevention of crop diseases.