Deep Learning-based Identification Of Silkworm Diseases

Silkworm diseases seriously affect the yield of silkworm cocoon,so automatic detection and timely elimination of silkworm diseases are particularly important in modern silkworm culture.In recent years,convolutional neural network technology has been developed rapidly and widely used in object detection.In this thesis,compared with several classical object detection networks,the YOLOv5 s network with the best performance was selected for suitability improvement,so as to improve the accuracy and speed of silkworm disease detection.By changing the backbone network and the Neck of YOLOv5 s,a new M-YOLOv5 s network model is designed.A new M-Resx residual structure was constructed by replacing the original standard convolution with a depthwise convolution with a 3×3 kernel,and then a new M-CSP feature extraction module was designed.CBH module is designed by replacing SiLu function with Hardswish activation function.M-CSP module and CBH module are used to replace the CSP module and CBS module of the original YOLOv5 s network respectively.In the Neck,GSConv convolution is used to replace the standard convolution to reduce the computational.In additoin,the coordinate attention is introduced in the last layer to improve the detection accuracy.To evaluate the impact of the above improvements on network performance,a data set of 5361 images of healthy and sick silkworms,including Beautarsia,pus,and piniosomiasis,was prepared.There are 3012 original images and the rest are obtained by image enhancement processing.The image enhancement techniques used include rotation,mirroring,brightness adjustment,Mixup and so on.Each image in the dataset was accurately labeled using the Labelimg tool with the advise of silkworm experts.The dataset was used to train,verify and test the network model.The results show that the designed M-Yolov5 s network has better comprehensive performance than the YOLOv5 s,the mAP is improved by 0.6%,the computation is decreased by 4.7 GFLOPs and the number of parameters is decreased by 2.31 M.Compared with the other nine classical networks,the experimental results show that M-YOLOv5 s model has the highest mAP,the smallest amount of computation and parameters,and the detection speed is only lower than YOLOv3-tiny network,up to 222 FPS.In addition,to meet the practical needs of industrial silkworm breeding,an intelligent silkworm disease detection software was developed,which can be deployed on the computer.The software can be run on Windows operating systems.It will display the classification and quantity of diseases as well as the marked images of diseases.It supports real-time detection of silkworm disease by input random pictures,recorded videos,or real-time videos.Users can adjust the intersection ratio and confidence to give different detection results.The software has high accuracy for silkworm disease detection,the detection results are intuitive and clear,and the operation interface is friendly.