Current methods for edge contour feature extraction for Unmanned Underwater Vehicle (UUV) side-scan sonar images have yet to solve the problems of low accuracy, discontinuous edges, and loss of detail. This paper proposes a new feature extraction method for UUV side-scan sonar images. By adding a batch normalization layer, the skip structure of the fully convolutional network (FCN) is improved, the redistribution of parameters in the skip structure is realized, and the training of the network is more sufficient. And we design a positive sample weighted loss function (WPSL) to improve the problem that the performance of the classification algorithm is degraded due to the imbalance of sample distribution in the data set. In this paper, an initial dataset is expanded by turning, rotating, and adding noise. An improved feature extraction network is then constructed, and the training of the improved FCN is completed by using a mini-batch gradient descent method, thus realizing accurate extraction of edge contour features of seabed topography. The experimental results show that the proposed method is more suitable to reject speckle noise than the traditional Canny and Fuzzy C-Means algorithms. Compared with current deep learning methods, the proposed method improves the ability to fuse detailed information and make discontinuous edges continuous. The mean intersection over union (IU) reaches 83.05%, which is 5.48% higher than the 77.57% before improvement.

当前的用于水下无人机（UUV）侧扫声纳图像的边缘轮廓特征提取的方法尚未解决精度低，边缘不连续以及细节损失的问题。提出了一种新的UUV侧扫声纳图像特征提取方法。通过添加批处理归一化层，改进了全卷积网络（FCN）的跳跃结构，实现了跳跃结构中参数的重新分配，并且对网络的训练更加充分。并且我们设计了一个正样本加权损失函数（WPSL），以解决由于数据集中样本分布不平衡而导致分类算法性能下降的问题。在本文中，通过旋转，旋转和添加噪声来扩展初始数据集。然后构造一个改进的特征提取网络，并通过使用小批量梯度下降法来完成对改进的FCN的训练，从而实现海床地形边缘轮廓特征的准确提取。实验结果表明，与传统的Canny和Fuzzy C-Means算法相比，该方法更适合抑制斑点噪声。与当前的深度学习方法相比，该方法提高了融合详细信息并使不连续边缘连续的能力。工会的平均交集（IU）达到83.05％，比改善之前的77.57％高5.48％。实验结果表明，与传统的Canny和Fuzzy C-Means算法相比，该方法更适合抑制斑点噪声。与当前的深度学习方法相比，该方法提高了融合详细信息并使不连续边缘连续的能力。工会的平均交集（IU）达到83.05％，比改善之前的77.57％高5.48％。实验结果表明，与传统的Canny和Fuzzy C-Means算法相比，该方法更适合抑制斑点噪声。与当前的深度学习方法相比，该方法提高了融合详细信息并使不连续边缘连续的能力。工会的平均交集（IU）达到83.05％，比改善之前的77.57％高5.48％。