改进PP-YOLOv2的水下侧扫声呐图像多目标识别

Multi-object Recognition of Underwater Side-scan Sonar Image Based on Improved PP-YOLOv2

针对水下侧扫声呐图像对比度低、噪声强度大,现有深度学习方法特征提取能力不足的问题,提出基于PP-YOLOv2引入注意力机制的改进侧扫声呐多目标识别方法.首先,针对侧扫声呐图像信噪比大、不同声呐设备生成的图像大小不一等特点,设计有效的图像预处理方法,包括噪声过滤、图像增强等;其次,基于当前目标检测性能很好的PP-YOLOv2模型设计改进,更换BotNet_dcn为模型主干网络,引入注意力机制提高网络特征提能力;最后,设计新的解耦头替换原耦合检测头,针对侧扫声呐图像的小目标进行精细化预测.结果表明:与原始PP-YOLOv2相比,所提方法在平均识别精度上提升了4.4%;与两种主流的基于卷积神经网络的方法相比,所提方法在平均识别精度上分别提升了4.66%和5.42%,同时在识别效率上分别提升32.4%和27.6%.

Because of the low contrast and high noise intensity of underwater side-scan sonar images,the feature extraction ability of existing deep learning methods is still insufficient.An improved side-scan sonar multi-target recognition method is proposed based on PP-YOLOv2 by introducing attention mechanism.First,for the characteristics of side-scan sonar images with high signal-to-noise ratio and different image sizes generated by different sonar devices,several effective image preprocessing method are explored,including noise filtering,image data augmentation,etc.Secondly,based on PP-YOLOv2,which is a state-of-the-art target detection method with good performance both in terms of precision and efficiency,a new model is designed by replacing the backbone network with BotNet-DCN.By doing this,the attention mechanism is introduced to improve the network feature improvement ability.Finally,a new decoupled head is designed to replace the original coupled head to perform refined prediction for small targets in sidescan sonar images.The results show that compared with the original PP-YOLOv2,the proposed method improves the average recognition accuracy by 4.4%;compared with the two mainstream methods based on convolutional neural network,the proposed method improves the average recognition accuracy by 4.66% and 5.42%,respectively,and improves the recognition efficiency by 32.4%and 27.6%,respectively.