Research On Apple Leaf Disease Image Classification Algorithm Based On Cross Supervision

As one of the origins of agriculture in the world,China has been a big agricultural country since ancient times,and the development of agriculture has always been one of the top priorities of the country.Apple is an important crop.Whether apple disease can be prevented or not directly affects the livelihood of many fruit farmers.Traditional apple leaf disease identification requires experts with relevant professional knowledge to inspect each fruit tree on site,which is time-consuming and inefficient.In the era of artificial intelligence,we can fully use machine learning to identify and classify apple tree diseases.Although various models for image classification have developed quite mature,training them mostly requires a large number of data samples.The real problem to be faced in the classification of apple leaf diseases is the difficulty in collecting these disease images.Therefore,there are few disease image samples in the real scene,and most of the data are raw data that has not been processed.Moreover,among various diseases of apple trees,there are more easily occurring diseases such as black rot,as well as relatively less easily occurring diseases such as cedar rust,resulting in category imbalance.Based on the above considerations,this paper uses Semi-Supervised Learning to solve the problem that it is difficult to obtain labeled samples in real scenes,and trains with unlabeled data when the number of labeled samples is insufficient.The research work of this article is based on the vertical research project " An Automatic Recognition Method for Apple Tree Disease and Pest Images Based on Metric Learning ".The research content of this paper comes from the task of "Model Construction and Algorithm Design" in this project.This article focuses on the classification task of apple leaf disease images,improves and enhances existing semi supervised work,and designs a new model.The main research content of the paper is as follows:1.In order to solve the problem that the semi-supervised Fix Match model tends to converge in the wrong direction due to the use of a single network,this paper designs a Cross Match model based on Fix Match by introducing cross pseudo-supervision.By using two networks with the same structure but different initialization and two data enhancement methods to combine and supervise each other,the problem that a single network is easy to produce the same wrong prediction and misguide the model is solved.2.In order to solve the problem of category imbalance caused by the scarcity of samples for certain disease categories or the difficulty of learning such disease characteristics,this paper introduces a gradient free prototype to design a PCM model based on Cross Match.Improve the model’s learning ability for difficult samples,and make the model learn the features of each category more evenly.3.In order to strengthen the ability of prototype to represent the disease features,so that the model can produce higher quality and quantity of false labels,this paper designs a hybrid cross-supervised model Blend Match based on prototype regularization according to the consistent regularization method.The linear branches supervise each other and the prototype branches,which improves the representation ability of the prototype and the prediction ability of the network.4.In order to verify the rationality and practicability of the above three models,experiments were carried out on CIFAR-10,CIFAR-100 and apple leaf disease leaves datasets,and compared with other semi-supervised classification models.At the same time,this paper also carried out some parameter experiments.The experimental results show that the semi-supervised model proposed in this paper can still obtain better performance under the condition of fewer apple disease samples,and has high practical application value.