A lightweight self-attention metric network for bird species recognition in intelligent bird repellent equipment

Abstract

Bird damages to power transmission lines pose significant operational risks, and intelligent bird repellent equipment (IBRE) requires accurate species recognition for effective long-term repellent. We propose a novel lightweight self-attention metric network (LSAM-Net) for few-shot bird species recognition in the vicinity of power transmission lines, aiming to enhance the performance of IBRE. LSAM-Net integrates a simple attention mechanism (SimAM) to emphasize critical spatial and channel features, thereby enhancing the extraction of key semantic information from bird images. Additionally, a self-correlation representation (SCR) module is employed to capture local structural patterns, effectively mitigating the impact of pseudo-features and improving the network’s capacity to learn discriminative representations. To promote the utilization of local discriminative information in few-shot classification, LSAM-Net leverages earth mover’s distance (EMD) to compute structural similarity between images. For efficient deployment, we apply knowledge distillation to further reduce model complexity. Extensive experiments conducted on Bird-65, CUB200, 2011, miniImageNet, and Fewshot-CIFAR100 demonstrate that LSAM-Net achieves superior performance compared to state-of-the-art methods, while maintaining a compact architecture. On the Bird-65 and CUB200-2011 datasets, LSAM-Net requires only 4.75 and 1.18 giga floating-point operations (GFLOPs), and achieves inference speed improvements of 52.9 % and 48.9 %, respectively, over the self-attention metric network (SAM-Net). Further optimization with TensorRT yields additional reductions in inference time by 43.6 ms and 53.7 ms, respectively. These improvements significantly support species-specific repellent strategies, thereby enhancing the long-term effectiveness of IBRE systems.