侧扫声纳检测沉船目标的轻量化DETR-YOLO法

Lightweight DETR-YOLO method for detecting shipwreck target in side-scan sonar

基于YOLOv5算法的侧扫声纳海底沉船目标检测方法虽然在检测精度和速度上取得了不错的成绩,但是如何在复杂海洋噪声背景下进一步提高小目标检测的准确性、降低重叠目标漏警和虚警率的同时实现模型的轻量化是一个亟需解决的课题。为此,本文创新融合DETR(end-to-end object detection with transformers)与YOLOv5结构,提出了基于DETR-YOLO模型的轻量化侧扫声纳沉船目标检测模型。首先,加入多尺度特征复融合模块,提高小目标检测能力。然后,融入注意力机制SENet(squeeze-and-excitation networks),强化对重要通道特征的敏感性。最后,采用加权融合框(weighted boxes fusion, WBF)策略,提升检测框的定位精度和置信度。实验结果表明,本文模型在测试集AP_0.5和AP_0.5∶0.95值分别达到84.5%和57.7%,较Transformer和YOLOv5a模型大幅度提高,以较小的效率损失和权重增加为代价取得了更高的检测精度,在提升全场景理解能力和小尺度重叠目标处理能力的同时满足轻量化工程部署需求。

Although the side-scan sonar shipwreck target detection method based on the YOLOv5 algorithm has achieved good results in detection accuracy and speed, however, how to further improve the accuracy of small target detection under the background of complex ocean noise, reduce the missed alarm rate and false alarm rate of overlapping targets, and realize the lightweight of the model is an urgent issue to be solved. To this end, this paper innovatively integrates the structure of DETR(end-to-end object detection with transformers) and YOLOv5,and proposes a lightweight side-scan sonar shipwreck detection model based on the DETR-YOLO model. Firstly, a multi-scale feature complex fusion module is added to improve the detection ability of small targets. Then, the attention mechanism SENet(squeeze-and-excitation networks) is integrated strengthen the sensitivity to important channel features. Finally WBF(weighted boxes fusion) weighted fusion frame strategy is adopted to improve the positioning accuracy and confidence of the detection frame. The experimental results show that the AP_0.5 and AP_0.5∶0.95 values of this model in the test set reach 84.5% and 57.7%, respectively, which are greatly improved compared with the Transfermer and YOLOv5 a models.At the expense of smaller detection efficiency loss and weight increase, higher detection accuracy has been achieved. At the same time, it can improve the full-scene understanding ability and the small-scale overlapping target processing ability while meeting the needs of lightweight engineering deployment.