Deep learning technology has been widely used in SAR ship detection tasks. However, complex sea level backgrounds, such as sea clutter and shorelines, greatly interfere with the accuracy of the detection of ship targets. In addition, embedded devices need to deploy multiple detection models, such as FPGA, so model size and detection speed are also important indicators in practical applications. In order to solve these problems, we developed a method combining traditional detection methods with deep learning. In this paper, SAR image clutter distribution model is used to suppress SAR image clutter, and then the processed image is sent to the network for learning. Based on this idea, we have established a superpixel composite Gamma distribution model, which can obtain more accurate fitting results than pixel scale Gamma distribution and suppress background clutter more effectively. We also propose the YOLO-SX lightweight detection network, which significantly reduces model size, detection time, calculation parameters, and memory consumption. Its overall performance is superior to other detection methods.