基于改进YOLOv8n的SAR舰船目标检测方法

A SAR Ship Target Detection Method Based on Improved YOLOv8n algorithm

摘要:

由于SAR图像舰船目标小、图像分辨率低、场景背景复杂等原因导致目标检测困难。针对上述问题，本文提出了一种基于改进YOLOv8n的SAR图像舰船目标检测方法。首先，将YOLOv8n的主干网络替换为CSWin Transformer，使网络能更好提取有用的细节特征。其次，设计了一种下采样弥补模块（Downsampling Compensate Module，DCM），弥补了下采样过程中可能丢失的有用信息。然后，设计了一种C2f-DWR(CSPLayer_2Conv-Dilation-Wise Residual)模块，增强了网络的多尺度特征提取能力。再者，嵌入了一种双动态token聚合器D-Mixer，赋予网络更强的归纳偏置能力和更大的有效感受野。最后，通过改进的混合域注意力机制，提高了模型的鲁棒性和泛化能力。SSDD数据集上的实验结果表明，与基线网络相比，本文方法在mAP@0.5和mAP@0.5:0.95性能上分别提高了1.7%和4.2%，显著提高了SAR舰船目标的检测性能。

Abstract:

Due to the small size of SAR ship targets, low image resolution and complex scene backgrounds, it is difficult to detect the ship targets in the SAR image. In order to solve the above-mentioned problems, a ship target detection method is proposed based on improved YOLOv8n is proposed. Firstly, the backbone network of YOLOv8n is replaced with CSWin Transformer, so that the network can better extract more useful detailed features. Secondly, a downsampling compensation module (DCM) is designed, which can retain more complete context information in the process of feature extraction. Then, a C2f-DWR (C2f-Dilation-Wise Residual) module is designed to enhance the multi-scale feature extraction capability of the proposed network. In addition, an embedded dual dynamic token aggregator D-Mixer is embedded, which can provide the network with stronger inductive bias ability and larger effective receptive field. Finally, the improved mixed-domain attention mechanism is embedded to improve the robustness and generalization ability of the proposed model. The experimental results on the SSDD dataset show that, compared with the baseline network, the proposed method improves the performance of mAP@0.5 and mAP@0.5:0.95 by 1.7% and 4.2%, respectively, which significantly improves the detection performance of SAR ship targets.