A deep learning-based high-precision and robustness target detection model for small crater detection

Deep learning-based methods for real-time small target detection are applicable to various industrial tasks, such as real-time traffic monitoring, military reconnaissance, and marine transportation. However, the high-precision detection of small objects in images with a complex background remains a challenge due to insufficient feature representation and background confusion. To address this challenge, this study proposes an efficient detector called multi-scale Feature Enhancement and Background Denoising YOLO (FEBD-YOLO), which includes a multiscale Feature Extraction Enhancement Module (FEEM) for effective feature extraction, a Hybrid Local Channel Attention module (HLCA) for enhancing foreground small target features, and a Background Denoising Module (BDM) for eliminating noise interference. The enhancement of the three modules in the feature extraction and feature fusion sessions provides the FEBD-YOLO with a high-precision detection capability for multi-scale small targets in images with a complex background. The proposed method can reach an accuracy of 0.782 (in terms of mAP@0.5) on the Martian crater datasets and 0.751 on the VEDAI datasets, surpassing several benchmark models and the other state-of-the-art models. In addition, the robustness of the FEBD-YOLO is verified under different simulated degradation conditions, which expected to be applied to satellite-based real-time detection tasks.