基于样本扩增的水下集装箱智能识别

An intelligent recognition of underwater containers based on sample augmentation

侧扫声纳是水下应急扫测的重要手段,水下目标识别是其中的关键技术,基于深度学习的智能识别技术现阶段因缺少大量样本效果不佳。基于侧扫声纳成像机理,针对拖鱼定位不准、船速不均匀和转向、声纳收发机制造成图像畸变问题,以集装箱为对象,提出一种基于弹性形变的样本扩增方法和目标识别方法。利用集装箱成像机理及畸变特点,首先对集装箱图像进行仿射变换;然后对该图像使用弹性变换(Elastic Distortions),随机调整形变控制参数生成指定数量的目标图像,完成样本扩增;在此基础上构建神经网络模型,实现侧扫声纳图像中的集装箱识别。试验表明,应用简单的弹性变换方法有效地扩增了样本,增强了神经网络的泛化能力,智能识别模型的查准率和查全率都提高了16%以上。

Side-scan sonar(SSS) is an important means of underwater emergency sweeping, and underwater target identification is one of the key technologies, and the intelligent identification technology based on deep learning is ineffective at this stage due to the lack of a large number of samples.Based on the imaging mechanism of side-scan sonar, this paper proposes a sample augmentation method and target identification method based on elastic deformation for the problem of inaccurate positioning of towed fish, uneven ship speed and steering, and image aberration from sonar transceiver manufacturing, taking container as the object.Using the container imaging mechanism and aberration characteristics, the container image is firstly affine transformed;then the image is randomly adjusted with Elastic Distortions to generate a specified number of target images to complete the sample augmentation;on this basis, a neural network model is constructed to realize the container recognition in side-scan sonar images.The test shows that the application of the simple Elastic Distortions method effectively expands the samples, enhances the generalization ability of the neural network, and increases the accuracy and completeness of the intelligent recognition model by more than 16%.