Research On Jujube Leaf Disease Identification Method Based On Deep Learning

Jujube cultivation is closely related to people’s daily needs and brings important benefits to the local economy of Xinjiang,but the leafy parts of the jujube tree are often subject to a variety of diseases during growth,due to the natural environment.Not only do these diseases seriously affect the quality and normal growth of dates,they can also cause huge economic losses to farmers.Therefore,the development of a model that can identify date palm leaf diseases is of great importance to improve the quality of red dates and reduce farmers’ economic losses.Based on this,this paper takes four common date palm leaf diseases(date palm white rot,date palm rust,mosaic leaf disease and date gall midge)as the research objects,shoots date palm leaf disease datasets under natural environment,carries out research in three parts: date palm leaf image pre-processing,date palm leaf disease recognition method and date palm leaf segmentation method,and constructs a date palm leaf disease recognition model.And the classification accuracy of date palm leaf diseases was improved.The main work of this paper is as follows:(1)This paper addresses the problems that manually collected datasets are prone to overfitting due to the small number of images compared to large publicly available datasets,and that the co-existence of four date palm leaf diseases and spot sizes makes detection and identification poor.In this paper,the dataset was expanded using horizontal,vertical and arbitrary angle random flip methods.The data set was expanded from 1934 to 9670 images,and the Labelimg and Labelme annotation tools were used to annotate the disease areas of the expanded date palm leaf disease images,generating TXT text and json text for the identification and segmentation data sets respectively.(2)To address the problems of disease spot occlusion and small disease targets in date palm leaf disease images with complex backgrounds that are prone to missed and false detection as well as low recognition accuracy,an improved YOLOv5 s model is proposed to call it the YOLOv5s-CBS model.By embedding the CBAM attention mechanism into the network prediction head,the network prediction head is thus enhanced for small target disease feature extraction,helping the model to better extract features from targets of interest,focus more on the disease targets of date palm leaves,and improve the problem of missed detection of obscured targets and small targets.The SIo U function is then used to speed up the model training and inference accuracy,and the feature fusion part of the backbone network is improved with a bi-directional weighted feature pyramid Bi FPN to enhance multi-scale feature fusion to improve the recognition effect at different scales.To better validate the effectiveness of YOLOv5s-CBS,it was compared with six models,namely Faster R-CNN,YOLOv5 m,YOLOv5l,YOLOv5 x,YOLOv5s and VGG16.Comparing the performance indicators,it was found that the best performance was achieved in the YOLOv5s-CBS model for date palm leaf diseases,with an accuracy of 89.23%,a recall of 91.46% and an m AP of78.86% at an orthogonal ratio of 0.5,with the highest accuracy of 97.9% and 94.7% for date white rot and date gall midge,respectively.In contrast,the identification of date rust and mosaic disease was relatively poor,reaching 67% and 55.5%,respectively.In summary,the model solved the problem of missed and false detection of date palm leaf diseases in the complex background with small target diseases,and improved the accuracy of date palm leaf disease identification..(3)A semantic segmentation-based method for date palm leaf disease recognition was proposed to address the problems of scattered and small features of date palm rust disease spots in early stage of date palm leaves and shallow spots of mosaic leaf disease and unclear spots against complex backgrounds.The segmented dataset was put into YOLOv5s-CBS for recognition,and finally the recognition accuracy of the dataset before and after segmentation was compared.The results showed that the recognition accuracy after segmentation was significantly higher than that before segmentation reaching 98%accuracy and 98%recall.m AP@0.5达到了99%The recognition accuracy of mosaic leaf disease and date rust improved by 42.5%and 24.3%,respectively,and that of date gall midge and date white rot improved by 4.6%and 1.5%,respectively.In summary,the method achieved the expected goal of improving the accuracy of date rust,foliar disease,date gall midge and date white rot identification,and is of reference value for practical application in date palm leaf disease identification methods.