<p>Accurate identification of Craters on the planetary bodies is essential for understanding the solar system’s evolution. Traditional crater detection methods often fail to accurately identify boundary results for complex craters on the lunar surface. Meanwhile, several deep learning (DL) architectures have been proposed for this problem, and a significant gap lies in assessing the performance of single-stage and two-stage object detectors by fusing the large datasets obtained from different planetary missions. To analyse the speed and accuracy of object detectors, this study evaluates several single-stage detectors (YOLO-v8, RetinaNet, and FCOS) and a two-stage detector (Faster R-CNN) using various extractors (DarkNet, Res2Net, ResNeXt, MobileNet, and ResNet) under different image resolutions. These techniques used transfer learning approaches to compare the efficiency of detectors, with a focus on the real-time scenario in this context. This study uses DEM, OHRC images, and Local DOM mosaics for experimental purposes, offering multi-source data with varying resolutions. This study demonstrates that two-stage detectors continue to provide the most reliable performance, despite the growing popularity of single-stage detectors. Faster R-CNN models are more accurate than single-stage detectors and are also more dependable when it comes to detecting multi-scale craters. Faster R-CNN with Res2Net-101 achieved the optimal accuracy, but in terms of training time, it takes longer than other object detectors; it still manages class imbalance problems efficiently. Furthermore, FCOS is marginally quicker than Retina-Net, with similar precision, but is not ideal for real-time applications in this context.</p>