Lightweight SAR Image Target Detection Algorithm Based on YOLO-v5

Abstract:
Synthetic Aperture Radar (SAR) has become one of the important means of marine ship monitoring and maritime traffic control because of its all-day and all-weather advantages. Therefore, the research on ship detection is of great significance. In traditional target detection, most scenes take the accuracy of detection results as the main purpose, while there is no high requirement for the size and complexity of the network model. To solve the above problems and maintain the necessary detection accuracy for the embedded platform, the algorithm proposed in this paper builds a lighter network model based on You Only Look Once version 5 (YOLO-v5). YOLO-v5-Light integrates the idea of MobileNet-v3, uses deep separable convolution to reduce the amount of network computation, and uses 1 × 1 convolution to fuse channels. The new network structure adds a bottleneck structure and a lightweight attention mechanism Squeeze-and-Excitation (SE), and changes the bottleneck structure to a spindle type. The parameters, model complexity and weight file size are compressed to 41.4%, 30.3% and 43.0% of the original network. The test experimental data are based on the SAR Ship Detection Dataset (SSDD), with a Recall rate of 91.4% and Mean Average Precision (mAP) of 88.7%.