基于连续帧与注意力机制的水下小目标自主检测算法

Underwater Small Target Detection Based on Continuous Frame and Attention Mechanism

摘要:

针对声呐对水下小目标探测时目标特征少，常规目标检测算法性能不佳的问题，提出了一种以 YOLOv5 目标识别算法为基础的连续帧识别改进算法。该算法通过连续帧数据提取模块以及轻量的通道空间注意力模块，提取声呐连续帧信息，提升了 YOLOv5 算法的识别能力。湖上前视声呐时序数据集算法验证试验表明，在几乎不增加推理时间的前提下，改进算法的平均检测精度比 YOLOv5 算法提升了 13.7%。该改进算法预期可在水下小目标自主检测任务中应用。

Abstract:

In order to solve the problems of few target features of underwater small targets collected by sonar， and poor performance of conventional target detection algorithm，an improved continuous frame recognition algorithm based on YOLOv5 is proposed. The algorithm uses a continuous frame data extraction module and a lightweight channel spatial attention module to extract sonar continuous frame information，which improves the recognition ability of YOLOv5 algorithm. The lake experimental results based on the forward-looking sonar time series dataset show that the accuracy of the algorithm is improved by 13.7%，and the reasoning time is basically unchanged. The improved algorithm is expected to be applied to the autonomous detection of underwater small targets.