Detection Method For Rape Pests Based On Deep Convolutional Neural Network

Rape is one of the important economic crops in China,and its output affects the development of China's national economy.Each year,rape pests have a great impact on the yield and quality of rape,posing a serious threat to the development of the oil industry.Therefore,accurate recognition and detection of rape pests is an important premise for effective control of rape pests.The traditional rape pest detection method is susceptible to background,illumination,angle,pest's posture and other factors,resulting in low robustness.Aiming at this problem,taking five kinds of common rape pests as the research object,this thesis proposed a rape pest detection model based on deep convolutional neural network.The experiment proves that the accuracy of the method is 94.12%,but there is a problem of inaccurate location and low recognition rate in small-scale rape pests detection.On this basis,this thesis proposed an improved detection method of rape pests,detecting the multi-scale rape pests accurately.The main research work of this thesis is as follows:(1)Based on the problem of poor robustness of currently available rape pest detection methods,a rape pest detection model based on Faster R-CNN was proposed.Experiments showed that this method had an accuracy of 2% higher than the traditional rape pest detection method,but there are certain difficulties in the detection of small-scale rape pests.(2)To improve the original model,this thesis proposes an improved rape pest detection model by modifying the network structure and increasing the small-scale anchor box,improving the detection performance of small-scale rape pests.In addition,hard negative example mining method is introduced to eliminate the imbalance between positive and negative samples,which enhances the discriminating ability of the model.(3)The thesis designed experiments,compareed and analyzed the detection performance before and after the model improvement.The effects of the number of anchor box,the number of convolution feature fusion layers and hard negative example mining on the performance of the improved rape pest detection model were also discussed.The experiment results showed that the average precision of the improved rape pest detection model was up to 95.01%,which was about 1 percentage point higher than before the improvement.The method can effectively detect multi-scale rape pests and show better generalization ability.Based on the improved detection method for rape pests,the problem of low robustness of traditional rape pest detection methods was solved,and the detection of rape pests can be more accurately implemented,which can effectively reduce the damage of rape pests and reduce the loss of rape.