Research On Crop Disease Recognition Based On Leaf Images

By using computer vision technology to identify crop diseases quickly and accurately is an important mean to ensure the harvest of agricultural products and promote agricultural modernization.This paper takes rape,cucumber,rice,corn,soybeans images as the research objects,doing some related research about disease recognition respectively under big samples condition and small samples.The main works are as follows:(1)In the stage of disease spots segmentation,considering the color characteristics of crop leaves,this paper firstly removes normal area of crop leave image under HSI color space,then merging it with the image which had been segmented by OTSU algorithm under Lab color space.Through the segmentation effects comparison with other different common algorithms,it shows this method can segment the disease spots from crop leaves more accurately.(2)In the stage of small-sample crops disease recognition,this paper takes the rape images as experimental object.Firstly color feature and texture feature are extracted.With the help of Euclidean distance,the basic probability assignment(BPA)which is necessary for D-S evidence theory can be constructed.Finally,using D-S combination rule of evidence to achieve the decision fusion and outputting the final recognition results through the decision-making conditions.In view of the situation that the final recognition result may be misrecognized as uncertain,this paper improves the fusion method by introducing the variance,which can avoid this defect.The experiment on the collected rape images obtains the recognition rate of 97.09%.The experiments show that this method can increase the rape disease recognition rate effectively.(3)In the stage of big-sample crops disease recognition,this paper selects rice,corn and soybean leaves as research objects.And Caffe tool is used to construct the continuous convolution layer convolutional neural network to extract the more advanced features.For Maxout unit is with better fitting ability,so it’s selected as the activation function.In order to reduce the parameters of the network,also to avoid over-fitting problem,Sparse Maxout unit is used in this network and effectively improves the performance of CNN.The experimental results show that the proposed algorithm is superior to the algorithms based on conventional CNN in the recognition of large crop diseases,and also better than the traditional manual extraction methods.