复杂SAR影像轻量级视觉感知模型

Lightweight Visual Perception Model in Complex SAR Image Scenarios

摘要:

复杂SAR影像场景下准确高效的目标检测算法能够提升监测装备的感知能力。针对海上复杂环境SAR图像舰船漏检、错检问题严重。提出一种复杂SAR影像场景下的轻量级视觉感知模型DH-YOLO（Double detection head-YOLO）。首先提出了低维双检测头网络结构，使用低维特征提取网络结构，融合双检测头完成特征融合和特征推理，解决SAR图像舰船小目标随着卷积导致的特征丢失问题；然后，提出了cosN-IoU损失函数，设计了新的锚框角度损失度量方式，并引入调节因子提高正样本的贡献。最后，提出了基于特征空间相似度的剪枝算法，依据滤波器相似度为可替换指标，完成网络特征信息剪枝，提高算法推理速度，降低算法复杂度、推理速度和模型体积，实现轻量化处理。为验证DH-YOLO算法的有效性，在SSDD、SAR-Ship-Dataset和HRSID中进行验证，结果表明，检测精度可分别提高至99.2%、98.5%、95.6%；融合剪枝算法，模型体积下降明显，可压缩至190KB；模型推理时间可降低至3ms以下。相较于当前主流算法，DH-YOLO在各方面均取得了不错的成绩，能够满足SAR图像实时舰船目标检测。

Abstract:

Accurate and efficient ship target detection algorithms can enhance the perception ability of monitoring equipment. There are serious issues of missed and false detections of ships in SAR images of complex marine environments. Propose a ship target detection algorithm DH-YOLO (Double detection head YOLO) based on YOLO for dual detection head SAR images. Firstly, a low dimensional dual detection head network structure was proposed, which uses a low dimensional feature extraction network structure to fuse dual detection heads for feature fusion and inference, solving the problem of feature loss caused by convolution of small ship targets in SAR images; Then, the cosN-IoU loss function was proposed, a new anchor box angle loss measurement method was designed, and an adjustment factor was introduced to improve the contribution of positive samples. Finally, a pruning algorithm based on feature space similarity was proposed, which uses filter similarity as a replaceable indicator to complete network feature information pruning, improve algorithm inference speed, reduce algorithm complexity, inference speed, and model volume, and achieve lightweight processing. To verify the effectiveness of the DH-YOLO algorithm, it was validated in SSDD, SAR Chip Dataset, and HRSID, and the results showed that the detection accuracy could be improved to 99.2%, 98.5%, and 95.6%, respectively; The fusion pruning algorithm significantly reduces the model volume and can be compressed to 190KB; The inference time of the model can be reduced to below 3ms. Compared to current mainstream algorithms, DH-YOLO has achieved good results in all aspects and can meet the requirements of real-time ship target detection in SAR images.