Oceanic eddies are ubiquitous phenomena carrying large amounts of energy, thus of great importance for marine ecology and sea-air exchange. As an all-weather, high-resolution sensor, Synthetic Aperture Radar could provide valuable observations for oceanic eddies. However, there are few wellestablished methods for eddy detection on SAR images except for manual seeking, which is laborious and time-consuming. In combination with deep learning, this study is among the earliest in the literature that attempts end-to-end eddy detection on SAR images. Due to obscure pictures and indistinct eddy boundaries, ordinary deep learning models are not adaptable to the objective. Therefore, an customized model, SANet-stacked attention network, is designed to recognize the unique eddy pattern presented on radar images automatically. SANet is a two-unit stacking architecture, with each an hourglass structure for bottom-up, top-down reference and the overall stacking network for iterative extractions of eddy textures contained in shallow layers of each unit. Besides, SANet has included the inner-hourglass attention gates and the outer-hourglass GCblock for the extracted features to be more concentrated on the interested areas. Using SANet, we have identified 87.75% of eddies in the constructed dataset collected from ESA-2 and ENVISAT SAR products. The result is much better than the no-stacking counterpart U-net, as well as state-of-the-art deep learning models DeepLabV3+ and SegFomer, thus verifying the superiority of the proposed method. The generalization ability of the algorithm has also been tested. The code and the constructed SAR dataset have been made public for broader use.