Research On Grain Insect Object Detection Algorithm Based On Deep Learning

Every year,millions of tons of food pests invade our country,causing huge losses to the country.Because it is impossible to find out the pest infestation in the warehouse in time,the fumigation method is often used to reduce the pest loss,but it also brings some environmental pollution problems.The food insect target detection method based on deep learning can predict the type and quantity of insect pests in the warehouse,which is conducive to taking targeted measures to reduce the loss in the warehouse.In this paper,a deep Convolutional NeuralNetwork(CNN)technique was used to detect five major beetle pests in the granary,and the high precision detection effect was achieved.The main research work of this paper is as follows:(1)Firstly,common methods for food and insect detection were studied.Domestic and foreign literature on food and insect detection was studied.The model,structure,algorithm and application evolution of convolutional neural network were analyzed.(2)Construction of data sets.By collecting images of five kinds of beetle food insects,the white board background food insect data sets and the actual stored food insect background food insect data sets were made,and the data enhancement method was used to expand the number of data sets,enrich the training data,and annotate and distribute the data in accordance with the Pascal Voc2007 format.(3)An improved Faster RCNN food insect target detection algorithm is proposed.The reasons for the low accuracy of Fathers RCNN algorithm in food insect target detection were analyzed,and the optimization strategy was proposed.Pyramid pooling module was added after the feature map to integrate the global context information.Focal Loss is used to replace the cross entropy Loss function as the classification Loss function to solve the imbalance problem of the data set.Experimental results showed that the improved Faster RCNN algorithm had an average detection accuracy of 89.42% and 90.12% for the food insect data sets of the white board background and the stored grain background,respectively.Compared with the previous improved algorithm,the improved Faster RCNN detection method proposed in this paper could better detect the stored grain pests.(4)The MF-SSD target detection algorithm was proposed.In view of the low recognition of food insect by SSD algorithm,the SSD algorithm was optimized,and the MF-SSD target detection framework was proposed to improve the missed detection and misdetection in food insect detection.In this method,VGG-16 backbone network was used for feature extraction,and MRFB module was designed to obtain richer context information based on the idea of adding receptive field to RFB module.In the part of multi-scale detection,the fusion mechanism of multi-scale feature layers is introduced to effectively fuse the low-level visual features and high-level semantic features in the network structure,and a residual prediction module is added to each feature layer for detection to further improve the model performance.The MF-SSD algorithm was tested on the white board background and the food insect data set of the actual stored grain background,and the MAP of the improved MF-SSD algorithm was increased by 2.63% and 8.68%,respectively.