Research On Pest Detection Method Based On CFFN Networ

The automatic detection of agricultural pests is of great significance to the monitoring,early warning and timely control of pests.The target detection method based on artificial intelligence can provide technical support for the automatic detection of agricultural pests.Because pest targets often exist in images as small targets and are accompanied by low feature resolution and weak carrying feature information,it is difficult to extract the pest feature information in the image,and the accuracy of pest detection is not high.Therefore,research on an efficient pest target detection method has important value for the realization of agricultural pest monitoring.Aiming at the problem that small pest images carry weak feature information and low resolution results in the insufficient ability of the network to extract feature information of small targets,this thesis proposes a pest detection method based on CFFN(Cross-level Feature Fusion Network)network.Faster R-CNN is selected as the pest detection network framework.First,in the feature extraction stage,a cross-level feature fusion structure CFF(Cross-level Feature Fusion)is designed,and the crosslevel feature fusion method is used to fully integrate different levels of feature information;Use the underlying feature map of the Res Net/Res Ne Xt backbone network as a reference image for information fusion to increase the accuracy of the feature information extracted by the network;The feature reconstruction module FR(Feature Reconstruction)is designed.By using sub-pixel convolution instead of upsampling,the feature information of the pests in the feature map is highly restored,and the aliasing effect of the CFF structure when fusing the feature information of different levels of feature maps is reduced.Secondly,during network training,design SOA(Small Object Augmentation)strategies to balance the area of small target pests and the background area,increase the effective mapping area of small pests in the CFF structure feature extraction process,and extract more small target feature information.The method in this thesis can fuse semantic feature information,visual feature information and location feature information of different levels of feature maps to generate feature maps for predicting small targets,which effectively improves the network’s feature extraction ability for small pests.Experiments show that the CFF structure and SOA enhancement strategy proposed in this thesis are effective,the detection accuracy rate reaches 88.1%,the recall rate and the F1＿score are 93.1% and92.3%,respectively.The experimental results verify the effectiveness of the method in this thesis.