Abstract: To address the challenge of balancing lightweight design and high accuracy in bird target recognition algorithms, we developed an improved model named YOLOv8-MAT-2H based on YOLOv8n, by fully considering the small size, sparse distribution, and complex backgrounds of birds in river and lake environments. First, we introduced the multi-scale feature module (MSBlock) to enhance the representation capability of bird features. Second, we adopted the adaptive down-sampling module (ADown) to strengthen extraction of edge and fine-grained features. Then, we employed a reduced detection head (Reduced Head) to trim the P5 branch, which significantly reduced parameter count and computational cost. Finally, we incorporated the adaptive-threshold focal loss (ATFL) into the loss function, so the model focused more on hard-to-detect targets. We conducted actual testing on a dataset containing 19003 images across 17 categories. The results showed that our model increased the mean value of average precision from 0.704 to 0.722, reduced parameters from 3.01 M to 2.41 M, and decreased giga floating-point operations per second from 8.1 to 7.3. Our approach could maintain real-time detection performance at 714.3 frames per second, and significantly enhanced responsiveness to small targets and complex backgrounds, providing an efficient and practical solution for deploying intelligent bird monitoring systems on edge devices in river and lake environments.

Key words: object detection, river and lake bird recognition, lightweight YOLOv8, data augmentation