| Agriculture is the foundation of my country’s national economic development.At present,agricultural pests and diseases greatly affect the yield and quality of crops.How to quickly identify and detect agricultural pests to achieve precise control is particularly important for improving the quality of agricultural production.Traditional pest classification and detection mainly rely on the naked eye of agricultural experts.However,there are many agricultural production areas in my country,the species of pests are complex,and there are differences within the same species,which leads to the low efficiency of this identification method and increases the cost of pest prediction and control.Therefore,the use of deep learning to automatically identify and detect pests and efficiently predict pest disasters has far-reaching significance.Based on the algorithm ideas and basic principles of deep learning,starting from the multi-scale feature extraction and channel attention mechanism,this paper uses the popular single-stage and two-stage instance segmentation models to segment agricultural pest images.The main work contents are as follows:(1)Pest image instance segmentation based on channel attention mechanism.Aiming at the problem of inaccurate segmentation of Mask R-CNN due to the similarities,different shapes and different sizes of pest species,the mask branch of Mask R-CNN model was improved based on attention mechanism and skip connection,and SE-Mask R-CNN pest instance segmentation model was proposed.First,the fivelayer network with the same size in the mask branch is changed to a seven-layer network,and the SE module is added to the first three-layer network to improve the feature extraction effect through channel feature weighting.Then,skip connections are made with the last three layers of networks respectively,and the fusion of multi-scale features is realized by combining context information.Experiments show that the improved algorithm has achieved better results on the agricultural pest data set constructed in this paper,which is 1.23% higher than the original network accuracy,reaching a recognition accuracy of 97.34%,which verifies the effectiveness of the branch proposed in this study.(2)Instance segmentation of pest images based on Res2 Net and FRe LU.Aiming at the problems of slow detection speed of Mask R-CNN model and low detection accuracy of single-stage network,Res2Net-YOLACT and Res2Net-YOLACT++algorithms based on Res2 Net module and FRe LU are proposed.Replacing residual blocks in the backbone network with the Res2 Net module can extract finer-grained multi-scale features and increase the receptive field range of the network layers.Using the FRe LU activation function to provide pixelated modeling capabilities can improve the accuracy of recognition and enhance the robustness of the model.The experimental results show that the improved Res2Net-YOLACT++ algorithm has better performance in the extraction of fine-grained features of agricultural pests.It achieves a classification accuracy of 91.72% on the agricultural pest dataset and has a high FPS of 27.7. |
