Giant Salamander Detection Model Based on Improved YOLO 11n-EWL

Abstract

Aiming to achieve rapid and accurate identification of giant salamanders in complex outdoor environments, an improved recognition model was proposed based on YOLO 11n. The model incorporated an efficient multi-scale attention module (EMA) in the backbone layer, replaced the complete intersection over union (CIoU) loss function with the Wise-IoU (WIoU) loss function, and introduced lightweight adaptive extraction of convolutions (LAE) in the head layer. Through ablation experiments and comparative tests, it was found that the improved model achieved 94.85% recall rate, 95.39% precision rate, 95.12% F1 score, and 77.20f/s frame rate, with a model memory footprint of 11.56MB, and the floating-point operation count was 8.65×109. Compared with the baseline YOLO 11n, the improved model outperformed it by 5.70 percentage points, 6.13 percentage points, 5.92 percentage points, and 27.1f/s in terms of recall rate, precision rate, F1 score, and frame rate, respectively. The proposed model YOLO 11n-EWL demonstrated significant improvements in model stability, recognition speed, and accuracy. The improved model can meet the real-time detection requirements of giant salamanders in the wild and can adapt to outdoor work in the long term, and can construct a set of giant salamander all-weather image intelligent recognition and behavior detection system. The research result can provide theoretical and technical support for the real-time detection of giant salamanders in complex outdoor environments.

Keywords: Andrias davidianus, object detection, YOLO 11, deep learning, giant salamander detection device