Research on multi class pests identification and detection based on fusion attention mechanism with Mask-RCNN-CBAM

This study addresses challenges in agricultural pest detection, such as false positives and missed detections in complex environments, by proposing an enhanced Mask-RCNN model integrated with a Convolutional Block Attention Module (CBAM). The framework combines three innovations: (1) a CBAM attention mechanism to amplify pest features while suppressing background noise; (2) a feature-enhanced pyramid network (FPN) for multi-scale feature fusion, enhancing small pest recognition; and (3) a dual-channel downsampling module to minimize detail loss during feature propagation. Evaluated on a dataset of 14,270 pest images from diverse Chinese agricultural regions (augmented to 7,000 samples and split into 6:1:3 training/validation/test sets), the model achieved precision, recall, and F1 scores of 95.91%, 95.21%, and 95.49%, respectively, outperforming ResNet, Faster-RCNN, and Mask-RCNN by up to 2.67% in key metrics. Ablation studies confirmed the CBAM module improved F1 by 5.5%, the FPN increased small-target recall by 6%, and the dual-channel downsampling boosted AP@50 by 3.1%. Despite its compact parameter size (63.87 MB, 1.39 MB lighter than Mask-RCNN), limitations include reduced accuracy in low-contrast scenarios (e.g., foggy fields) and GPU dependency. Future work will focus on lightweight deployment for edge devices and domain adaptation, offering a robust solution for intelligent pest monitoring systems that balance accuracy with computational efficiency.