Abstract: High-precision and efficient automatic crater detection is crucial for planetary geological research, surface age estimation, and rover autonomous navigation. However, deep learning-based detection methods often suffer from insufficient feature representation when addressing challenges such as blurred crater edges and huge scale variations in planetary remote sensing images. To tackle these issues, this paper proposes a crater detection method for Martian thermal infrared remote sensing images. A Sobel edge-guided attention (SEGA) module is designed to alleviate the problems of blurred edges and degraded structures of craters. By introducing ring-shaped edge prior information based on the Sobel operator into the backbone network and combining it with a cross-attention mechanism, the network is guided to focus on the contour and structural features of craters. To address the large-scale span of craters and insufficient feature expression, a wavelet feature fusion module (WFM) is constructed. By integrating multi-resolution frequency-domain information extracted via wavelet transform with a channel attention mechanism, the module enhances the feature fusion capability for craters and effectively copes with extreme scale variations from mini-craters to large impact craters. Experimental results on the public Mars Day Crater Detection dataset demonstrate that, compared with the original YOLO series models and other mainstream detection models, the proposed AW-YOLO (you only look once with attention and wavelet) model achieves favorable performance while maintaining high inference efficiency. Specifically, its Precision, Recall and F1-score reach 0.88, 0.87 and 0.87, respectively.