Pest recognition algorithm based on attention mechanism

Accurate identification of crop pests is of great significance for timely protection and management of crops. When facing the natural environment, due to the small size of crop pests, little difference in color from the environment, and the influence of various noise and complex background factors, the current crop pest recognition model related to deep learning is difficult to meet the requirements of recognition accuracy and robustness, and it is difficult to deploy on mobile terminals with limited computing resources and low performance. Therefore, ShuffleNetV2 0.5×, which has the least model parameters in the ShuffleNetV2 network structure, is selected as the benchmark network to design a lightweight crop pest recognition model (Shuff-CANet) based on attention mechanism. Among them, coordinate attention (CA) mechanism can further suppress background noise and focus on key information of crop pests, and effectively enhance the robustness of the model. Bistable logic loss function with label smoothing regularization (LSR) can solve the two shortcomings of logic loss function in training noisy data sets, and make the model more adaptive to noise. The experimental results show that the Shuffle-CANet model achieves 90.42% recognition accuracy on the test data set of 10 common crop pests in natural scenes, and the recognition accuracy is higher than the existing classical convolutional neural networks VGG-16, GoogLeNet, Xception, etc. The model has the characteristics of high recognition accuracy, strong robustness and good anti-interference performance, and can be well adapted to the practical application of crop pest recognition.