Crater-DETR: A Novel Transformer Network for Crater Detection Based on Dense Supervision and Multiscale Fusion

The crater detection (CD) presents challenges due to complex backgrounds, tiny scales, and dense distribution of small craters. To address these problems, we propose a new DEtection TRansformer (DETR) variant network for CD called Crater-DETR. First, we design the correspond regional attention upsample (CRAU) and pooling (CRAP) operators through cross-attention computing to address the problem of foreground–background confusion caused by the feature loss of small craters. Then, to address the weak supervision due to the fixed number query selection and to introduce dense supervision, we propose the dense auxiliary head supervise (DAHS) training. Next, automatic denoising (ADN) training is proposed to solve the problem of sparse positive queries in the decoder to improve the decoding capability. Finally, we propose a small object stable intersection over union (SOSIoU) loss to optimize the training process since the matching process is more unstable in small craters compared with other sizes of craters. Experimental results show that Crater-DETR achieves a precision of 88.13% on the DACD dataset, yielding state-of-the-art performance. Furthermore, our method also achieved the best performance on the ISDD and AI-TOD datasets, verifying its robustness.