| Pests are a significant factor affecting rice growth and yield.To improve rice yield,pest classification and trend prediction are necessary,and targeted pest control measures need to be provided.Traditional pest classification is usually manually performed,but this method is inefficient,inaccurate,and unable to meet practical production needs.Therefore,this study proposes a rice pest classification method based on the BCNN network,with the main research content including:(1)Constructing a pest image dataset.The images in the dataset were obtained through pest monitoring devices and consist of 1647 images.By applying data augmentation operations,the number of images was increased to 11143,resulting in a self-made rice pest image dataset that can be used for related research on rice pest classification.(2)To address the problem of missed or incorrect classification of small-target pests,an optimized BCNN algorithm is proposed.Firstly,the shallow second-order features of the BCNN network are obtained by introducing bilinear convolution operations to enhance the shallow feature extraction capability of the network.Secondly,deep convolution operations are performed on the shallow bilinear features to obtain second-order deep features,improving the discrimination ability of the network for the deep semantic features of pests.Finally,the second-order deep features are up-sampled and fused with the shallow features to obtain the MF-BCNN model,which improves the multi-scale pest classification performance.(3)To address the issue of insufficient inter-layer information interaction in the convolutional layers leading to a decrease in model performance,the MF-BCNN algorithm is further optimized.Firstly,the axial attention mechanism is introduced into the MF-BCNN network model to compensate for the lack of network feature information interaction and to enhance the network’s learning and generalization capabilities.Secondly,the loss function is changed to the Large-Margin Softmax to improve the performance of the fine-grained pest classification model.Finally,the model is trained to obtain a network model suitable for rice pest classification,with a classification accuracy of 96.56%.(4)To facilitate user operation of the pest classification model,a pest intelligent monitoring system based on a web platform was developed,with HTML5 used to write relevant pages,including classification recognition,statistical analysis,pest warning,and user management pages,providing effective support and guidance for rice pest control. |
