Edge Deployment of SAR Ship Detection Network Based on Lossless Model Compression and Quantization-Aware Training

‍ ‍Methods based on deep neural networks have shown great advantages in the task of detecting ship targets in synthetic aperture radar （SAR） images. However， the large number of parameters and computing power requirements make it difficult to deploy in edge environments with limited resources. To address this problem， we improved the single-stage target detection network known as You Only Look Once （YOLO） v5s from two aspects： network lightweighting and model deployment optimization， and we propose a SAR image ship target detection network deployment method for edge environments. In terms of network lightweighting， we combined channel-level network pruning based on scaling factors of batch normalization layer and fine-grained knowledge distillation based on feature responses to achieve lossless compression of the ship-detection network. The parameter amount and calculation amount of the lightweight model decreased by 80.3% and 51.3%， respectively， compared with the baseline， without causing a loss in detection accuracy. The average accuracy on the SAR ship detection dataset was 0.979 （the baseline was 0.980）. For model deployment optimization， we propose a mixed-precision tensor real-time （TensorRT） inference engine guided by quantization-aware training based on an embedded GPU， which greatly improved the model inference speed and reduced the operating power consumption of the device. The inference speed of the lightweight inference engine on an SAR image with a size of 640 × 640 pixels was 208 frames per second， reaching 3.41 times that of the baseline. Simultaneously， the inference power consumption of the device was only 6.2 W， which was a 61.0% decrease compared to that of the baseline. In addition， benefiting from quantization-aware training， the mixed-precision TensorRT inference engine achieved similar inference speed and power consumption as the 8-bit integer precision TensorRT inference engine； however， with an increase of 44.1% in the average accuracy， which was only 0.9% lower than the baseline value. Experimental data showed that the method proposed in this article can well take into account the requirements of real-time measurements， accuracy， and low power consumption for ship target detection in SAR images in edge environments.