An improved RT-DETRResnet18 algorithm for wild animal recognition on the Qinghai-Tibetan Plateau

Abstract In response to the unique geographical and ecological environment of the Qinghai-Tibet Plateau and its distinctive biodiversity, this study proposes an optimized algorithm, RT-DETR-PATD (Plateau Animal Target Detection), based on RT-DETRResnet18, for better protecting the wildlife on the Qinghai-Tibet Plateau. The model employs a feature pyramid structure to receive information from three scales for feature reconstruction. Through spatial feature reconstruction, dynamic interpolation, and multi-feature fusion, the model enhances its multi-scale feature representation and improves its ability to recognize targets in complex backgrounds. Additionally, the model incorporates an Adaptive Fine-Grained Channel Attention (AFGC) module, which significantly increases detection accuracy. The improved model has increased its precision from an original 79.6% to 82.7% on mAP50 and from 61.6% to 63.9% on mAP50-95. This model demonstrates high accuracy in our Qinghai-Tibet Plateau wildlife dataset, meeting the requirements for wildlife identification on the Qinghai-Tibet Plateau.