A Computer Vision-Based Detection Model for Wild Chinese Giant Salamanders in Complex Environments

To achieve rapid and accurate identification of wild Chinese giant salamanders in complex field environments, we propose an improved recognition model based on YOLO v11n. This model incorporates an Efficient Multi-scale Attention Module (EMA) into the Backbone layer, replaces the Complete Intersection over Union (CIoU) loss function with Wise-IoU (WIoU) loss, and introduces Lightweight Adaptive Extraction of Convolutions (LAE) into the Head layer. Ablation and comparative experiments demonstrate that the improved model achieves recall, precision, F1 score, and frame rate of 94.85%, 95.39%, 95.12%, and 77.20 f/s, respectively. The model occupies 11.56 MB of memory and performs 8.65×10⁹ floating-point operations. Compared to the baseline YOLO v11n, the recall, precision, F1 score, and frame rate are 5.70, 6.13, 5.92 percentage points higher and 27.1 fps faster, respectively. The proposed YOLO v11n-EWL model demonstrates significant improvements in stability, recognition speed, and accuracy. The improved model meets the real-time detection requirements for wild Chinese giant salamanders in natural habitats and can sustain long-term outdoor operation. Based on this, an all-weather image recognition and behavior detection system for wild salamanders was developed, successfully detecting over 20 instances of wild Chinese giant salamanders. This system provides a stable and reliable monitoring tool for advancing the conservation of wild salamanders and their habitats.