Research On Intelligent Processing Of Peach Disease Images Based On Deep Learning

Plant diseases severely affect the normal growth of plants,resulting in a decrease in yield and economic income,and even affect people’s dietary health.Therefore,only by identifying the type of plant disease in time and displaying the focus area and distinguishing the severity of the disease can effective measures be taken for treatment and prevention.Due to its powerful performance,deep learning is widely used in the research of plant diseases.However,there are less studies on peach tree diseases in the use of deep learning methods.Therefore,this paper conducts research on a self-made peach disease image data set.The research content has the following three parts:1.Deep learning was used to classify peach tree diseases.In the self-collection peach tree disease dataset,the samples is not enough,and the sample of different diseases is not the same.In the case of fewer samples and unbalanced samples,the classification accuracy of the deep network model will be lower,so this paper proposes two methods to solve this two problems.One method is using Xception + L2/L2 M regularization.L2 M regularization is an innovative regularization expression proposed on the basis of L2 regularization,which mainly solves the problem of low classification accuracy caused by fewer samples;The other method is to weight the sample,which mainly solves the problem of low accuracy caused by imbalance sample.Experiments show that both methods have great improvements compared to the original model.2.Deep network models were used to segment peach tree diseases.The current literature uses a deep network model to segment plant diseases,which is semantic segmentation.Although the diseased area can be displayed well,people prefer to get the name of the area and disease and mark of the segmentation on the original image.Instance segmentation can can solve this problem.Mask R-CNN and Mask Scoring R-CNN network models are both models for instance segmentation.But one step in the network is to train the classification of foreground and background samples,and the samples of foreground and background samples are unbalance and hard.If the classification accuracy at this stage is not high,it will directly affect the final segmentation results.this article proposes using Focal Loss to slove this problem at this stage.Focal Loss can solve the problem of reduced accuracy caused by hard samples and by unbalanced samples.The experimental results show that this method can effectively improve the segmentation effect.3.Instance segmentation model was used to evaluate the disease level of peach diseases.When the current literature uses the deep network model to evaluate the plant disease level,most of the samples are marked by the level of disease classification,and then the deep network model is used for classification.The output is the level of the disease.This method cannot visually see the damaged area of the picture.Based on the appearance and application of instance segmentation,this paper uses the instance segmentation method to evaluate the peach disease level.The method firstly marks the disease level of peach disease samples,and then uses the improved Mask R-CNN network to output the disease name,disease level results and the Mask of the lesion area at the same time.Since the segmentation effect of the Mask R-CNN network is not well,this paper changes the segmentation branch to the Refine Mask branch to improve the segmentation effect.In addition,in the backbone network,the Res Net module is changed to a multi-scale Res2 Net module to extract features in more detail.The effect is also improved.In this paper,a image dataset of peach diseases was established.First,the loss function optimization was used to improve the classification accuracy of peach diseases in the case of unbalanced samples or few samples,and then Focal Loss based on instance segmentation was used for the situation that the foreground and background of the samples are difficult to classify.The segmentation method was used to segment the disease area.Finally,the improved instance segmentation was used to evaluate the disease level and obtain the segmentation results at the same time.The research in this paper provides support for obtaining the effective and detailed information of peach diseases.