RHAM-CA-Det: An Attention Mechanism for Butterfly Detection

Deep learning-based butterfly detection plays a crucial role in biodiversity research. In practical applications, the complexity of field environment and variety of butterfly morphologies (such as size, color, etc.) present significant challenges to computer vision algorithms. To address these challenges, we propose RHAM-CA-Det, an object detection model designed for robust butterfly detection. Based on Faster R-CNN, we introduce Channel Attention with Learnable Integration (CALI) to help model focus on the most informative regions and Regional Hybrid Attention Module (RHAM) to apply the self-attention mechanism for the important regions, while using convolution for insignificance regions. Extensive experiments on a public butterfly dataset demonstrate the superior performance of our model. As a result, the proposed RHAM-CA-Det achieves 68.1% mAP, surpasses both the mainstream model YOLOv8 and baseline by 9.0% and 4.4%, respectively, demonstrating potential for butterfly detection applications. The code is available at our GitHub repository.