Research On Detection Of Silkworm Nuclear Polyhedrosis Based On Convolutional Neural Networks And Development Of Early Warning Software

Nuclear polyhedrosis is a silkworm disease that has a high frequency and infectiousness during the instar stages in commercial rearing,and is the main factor for the reduction of silkworm cocoons in China.Current nuclear polyhedrosis prevention and control methods are mainly the selection and breeding of silkworm varieties that are resisitant to the virus that causes nuclear polyhedrosis,the use of large amounts of disinfectant to thorougly clean silkworm housing and tools,and spreading lime of powder in silkworm rearing beds during feeding to inhibit the growth of pathogens.Manual inspection,early warning and prevention are relied on when the disease spreads.However,the current methods of disease control and early warning are time-consuming,have a limited scope of application,and are only suitable for traditional manual silkworm rearing.As sericulture industry gradually upgrades to mechanized breeding,there is an urgent need for a new method of early warning and prevention of nuclear polyhedrosis.The aim of this study was to investigate a new method for addressing this issue,according to silkworms infected with nuclear polyhedrosis will show different visual characteristics with healthy ones in same growth stage,and the excellent performance of deep learning technology in field of vision,convolutional neural networks were used to detect images of healthy silkworms and those with nuclear polyhedrosis,and could provide a new method for early warning and prevention of silkworm nuclear polyhedrosis.The main work completed includes:(1)An image dataset for detection of silkworm nuclear polyhedrosis was constructed.The methods of breeding and infecting pathogens were used to collect both healthy and diseased samples of the silkworm variety Chuanshan×Shushui at several growth stages.The placement of diseased and healthy silkworms in background was used to simulate the early stage of diseases spread which needs timely warning and prevention,and 941 original images were taken in natural environment to construct silkworm nuclear polyhedrosis detection image dataset.An image tool was used to label categories and positions of all silkworms in the dataset.A total of 1493 healthy silkworms and 1479 silkworms with nuclear polyhedrosis were labeled.(2)A detection algorithm for silkworm nuclear polyhedrosis was designed.The basic structure and operation flow of convolutional neural networks were described.The silkworm nuclear polyhedrosis algorithm was designed by reducing network structure of YOLO v3 that representative algorithm of one step detection principle to improve the operating efficiency of the network,based on the analysis of dataset in this article and experiments.The network structure,data encoding method,prediction result decoding and screening process of the reduction algorithm were described.The K-Means algorithm was used to cluster position boxes information of silkworms in dataset,and anchor boxes suitable for dataset in this study were obtained.(3)Detection tests of silkworm nuclear polyhedrosis were carried out.A visual experiment of convolution operation was carried out,and the results showed that convolutional neural networks can automatically extract features of silkworm images in complex background,and iteration can effectively enhance the feature extraction ability of the network.The mean average precision(m AP)and detection time of networks on testset were selected as the evaluation index,a test of the hyperparameters’ impact on detection was carried out,and the results showed that different hyperparameters have a varying impacts on detection results,and when learning rate is 0.001(dynamic),the optimizer is Adaptive Moment Estimation(Adam)and a small batch＿size,the convergence performance and detection effect of network are the best.A test of influence of parameter layers on detection was carried out.18 networks with different parameter layers were constructed by using YOLO v3 as the basic frameworks and gradually increasing the number of parameter layers for testing.The results showed that adding parameter layers within a certain range can improve mean average precision,but on the same parameter layer,too many convolution kernels would cause a significant increase in the amount of training parameters,and may also lead to a decrease in mean average precision.A comparison test between the reduction algorithm and standard YOLO v3 was carried out,and the results showed that the modified method in this study can improve the calculation speed and detection accuracy of the reduction network,with mean average precision of 94.77%,and average detection time of a single image is 0.031 s,which are higher than standard YOLO v3 algorithm.(4)Detection and early warning software for silkworm nuclear polyhedrosis was developed.The software function requirements were analyzed.Software interface and function program were designed,and the reduced YOLO v3 network and pretrained weights file were deployed to early warning software.The software has three working modes,which can respectively detect real-time video of camera,local video and pictures,and automatically display warning information when it detects silkworm with nuclear polyhedrosis.The software can also save detected images and result text,which is convenient for expanding dataset,and optimizing and improving detection algorithm.The results of this study showed that convolutional neural networks can achieve efficient and accurate detection images of healthy silkworms and those with nuclear polyhedrosis.It can provide a new method for disease early warning and control for mechanized silkworm rearing mothods,and have very broad application prospects for sericulture industry.