Metric learning unveiling disparities: A novel approach to recognize false trigger images in wildlife monitoring

Wildlife monitoring using camera traps is a vital tool for ecosystem health assessment. However, camera traps often face high rates of false-triggered images (empty shots), significantly impacting data processing efficiency. This study proposes a metric learning-based method for false-triggered image recognition. By integrating K-means clustering for sample selection and a triplet loss function for model optimization, the approach effectively distinguishes subtle feature differences in false-triggered images. Experiments demonstrate that the proposed method achieves 80.17% Accuracy, 79.79% Recall, and a reduced false positive rate (FPR) of 19.48% on test datasets collected from various regions. Compared to traditional models, it improves Accuracy and Recall by 5.5% and 5.96%, respectively, while reducing the FPR by 5%. On embedded device Jetson Nano, the method achieves a single-image inference time of just 0.076 s, showcasing its potential for deployment in resource-constrained environments. This research addresses challenges related to high intra-class diversity and inter-class similarity in false-triggered images, offering a novel solution to enhance wildlife monitoring efficiency. The code is available at https://github.com/hzl-bjfu/AIPL/tree/master/RFTI . • A metric learning-based approach for false-triggered image recognition is proposed, significantly enhancing feature distinction. • The method leverages K-means clustering to improve training efficiency and sample diversity. • Achieves single-image inference in 0.076 s on embedded devices, making it suitable for resource-constrained environments.