Abstract: In order to investigate the performance of different deep learning object detection algorithms in submesoscale and small-scale ocean eddy detection, this study constructs a SAR image dataset containing submesoscale and small-scale ocean eddies using satellite images such as ERS-1/2, ENVISAT, Sentinel, GF-3 and ALOS-2, which covers multi-source and multi-scale ocean eddy targets. Based on the constructed dataset, three object detection networks, RetinaNet, Faster R-CNN and Cascade R-CNN, are used to conduct experiments, and a comprehensive comparative analysis is conducted to evaluate the detection speed, accuracy and background clutter resistance of three networks. Experimental results indicate that RetinaNet is the fastest in terms of detection speed, achieving a frame rate of 19.4 frames per second. For detection accuracy, Faster R-CNN outperforms RetinaNet and Cascade R-CNN with an average precision of 0.797 5. When it comes to background clutter resistance, Cascade R-CNN demonstrates greater strength, excelling at detecting eddies in densely distributed and less distinguishable backgrounds.