Cotton Field Pest Recognition Based on Gradient Descent Mechanism and Large Convolutional Kernel Network

To assist farmers in efficiently managing cotton fields in industrial areas, we propose a pest image recognition solution. This solution is based on the GD Mechanism (Gather and Disturb Mechanism) and a large convolutional kernel network. It incorporates the FAM (Feature Alignment Module), IFM (Information Fusion Module), and Inject (Information Injection Module) to align, fuse, and inject features at different hierarchical levels. This enhances the feature fusion capability of the network structure. To boost the model's sensitivity to enlarge the receptive field, we employ the RepLKNet with a large convolutional kernel as an enhancement to the existing small convolutional kernel in the original CNN architecture. Experimental results indicate a significant improvement in recognition accuracy after modifying the model. The optimized model's average precision increases by 2.8%, reaching 96.2%. This provides strong support for preventing and identifying cotton pests at different stages.