Research on Multi Class Pest Identification and Detection Based on Fusion Attention Mechanism with Mask-RCNN-CBAM

Xingwang Wang¹Can Hu²Xufeng Wang²Hainie Zha^3*Xueyong Chen^1*Shanshan Yuan⁴Jing Zhang⁵Jianfeng Liao⁵

• ¹College of Mechanical and Electronic Engineering, Fujian Agriculture and Forestry University, Fuzhou, China
• ²College of Mechanicaland Electrical Engineering, Tarim University, Alar, China
• ³University Key Laboratory of Intelligent Perception and Computing of Anhui Province, Anqing Normal University,, Anqing, Anhui Province, China
• ⁴Quality and Safety Inspection and Testing Center for Agricultural Products, Qiemu County, China
• ⁵Anhui Eagle Information Technology Co., Ltd, Anqing, Anhui Province, China

With the advancement of agricultural modernization, pest identification and extraction have become increasingly pivotal in crop pest control. To address the issues of false positives and missed detections caused by complex backgrounds and densely packed pest aggregation, this paper proposes an enhanced Mask-RCNN model integrated with a Convolutional Block Attention Module (CBAM) for precise pest extraction in complex environments. In particular, this research focuses on mitigating the missed and false detections that commonly arise in densely populated pest regions by introducing the CBAM attention mechanism, a feature-enhanced pyramid network (FPN), and a dual-channel down sampling module into the original Mask-RCNN architecture. The incorporation of the CBAM attention mechanism allows the model to focus more effectively on the detailed features of the pests while minimizing the interference from background noise. The feature-enhanced FPN module facilitates the fusion of multi-scale feature maps, improving the detection capability for small target pests. Additionally, the dual channel down sampling module optimizes the information retention during feature propagation, further enhancing the model's overall performance. Experimental results demonstrated that the proposed model outperforms existing algorithms such as ResNet, Faster-RCNN, and the original Mask-RCNN in terms of precision, recall, and F1 score, with improvements of 0.73%, 1.38%, and 2.67%, respectively. Moreover, the method significantly reduces false positives and missed detections, achieving a favorable trade-off between computational efficiency and model size, with a reduction of 1.39 MB in parameters compared to the original Mask-RCNN. Ablation studies further substantiate the contributions of the attention mechanism, feature enhanced FPN, and dual channel down sampling modules to the performance gains. Overall, the enhanced Mask-RCNN-CBAM model demonstrates superior performance in complex backgrounds and dense pest regions, exhibiting substantial practical value and promising application prospects.