Research On Image Recognition Of Crop Leaf Disease Based On Transfer Learning

Two basic assumptions should be met when using the traditional machine learning methods to identify the disease images of crop leaf : 1)the training samples and the test samples should meet the conditions of independent and identical distribution;2)sufficient labeled samples are needed to obtain an ideal classification model.In practical application,it is difficult to collect massive of labeled images for a certain task: the occurrence of crop leaf diseases is affected by season,region and climate,which increase the difficulty of image collection;marking a large number of sample images takes manpower and material resources.In recent years,more and more experts and scholars began to use transfer learning to solve the problem of insufficient labeled data in the target area,and improve the utilization rate of the existing resources,further improve the learning effect of the tasks in the new fields.During the process of crop leaf disease image recognition,there appear the problem of class imbalance and other problems in the data sets.Instance based transfer learning and model based transfer learning were studied in this paper.The main research contents are as follows:1)In order to solve the problem of class imbalance in the target data sets,an improved image recognition method of crop leaf diseases based on the improved Tr Adaboost algorithm was proposed by combining the ensemble learning with the instance based transfer learning.The image preprocessing method named whitening was used to reduce the image redundancy and decrease the influence of low image classification caused by exposure,and what's more C-means clustering method was used to obtain images in the source domain data sets which were similar to these in the target data sets.In order to improve the recognition rate of the minority classes in the target data sets,a dynamic balance factor was added to the minority class weights of the target data sets.The weights of the minority classes were updated according to the error rate of the majority class and the minority class,and the traditional classifier is replaced by support vector machine.The experimental results showed that the proposed method can improve the recognition rate of minority classes in the target data sets.2)In order to solve the problem that the feature complexity of corn leaf disease image was so high,a method of corn leaf disease image recognition based on deep transfer learning and metric learning was proposed.The Alex Net was selected to extract image features,andmeasurement layer was added in front of the classifier.The average Euclidean distance between the generalization center of each type of samples was calculated as the Center Loss,then the loss function and the classifier loss were used as the total loss function of the network to guide the network training,so as to increase the distances of inter-class and reduce the distances of intra-class.In addition,a hard training strategy was applied to improve the accuracy of hard training samples recognition,and further more improve the speed of network training.The experimental results showed that the proposed method was effective in the recognition of corn leaf diseases images with high complex characteristics,as well as the high convergence speed.3)Aiming at the problem that the number of labeled samples in the target domain was small as well as the feature distribution in the target domain was quite different from that in the source domain,an image recognition method based on the depth adaptive network was proposed.In order to extract the image features of the source domain and the target domain more effectively,a two-branch Alex Net with shallow network parameters sharing was used to extract the image features of the source domain and the target domain respectively;adaptation layers were added to the last three fully connected layers of the network,and the Maximum Mean Discrepancy was introduced to measure the feature distribution discrepancy between the source domain and the target domain,so as to achieve the goal of the target domain classification by reducing the feature distribution discrepancy.The experimental results showed that the proposed method can improve the recognition achievement of the target domain when features distribution of the source domain was quite different from that of the target domain.