【摘要】 为了提高基于侧扫声纳图像提取海底沙波谷线这种类别不均衡线状地物的精度，提出了一种深度学习与数学形态学相结合的方法。该方法采用Dice损失函数和添加批标准化(batch normalization, BN),对U型卷积神经网络模型(U-Net)进行改进；结合数学形态学中的闭运算和骨架法，对沙波谷线轮廓进行修复并提取线性特征；进一步将改进的U-Net模型与支持向量机(support vector machine, SVM)、随机森林(random forest, RF)、面向对象分类以及U-Net模型进行精度对比验证。结果表明：改进的U-Net模型能够解决类别不均衡的问题，实现沙波谷线的高精度提取，该方法对海底沙波的研究具有重要的科学与工程应用价值。更多还原

【Abstract】 In order to improve the precision of submarine sand wave trough lines extracting from side scan-sonar image, such as this kinds of unbalanced linear features, a new method combining deep learning and mathematical morphology was proposed. The U-shape convolutional neural network model(U-Net) was modified by using Dice loss function and adding batch normalization(BN). In combination with closed operation and skeleton method from mathematical morphology, the contours of the sand wave trough lines were repaired and linear features extracted. Furthermore, the accuracy of the modified U-Net model was compared with support vector machine, random forest, object-oriented classification and U-Net method. The results showed that the modified U-Net model can solve the problem of class imbalance and achieve high-precision in extracting the sand wave trough lines. The proposed method has significant scientific and engineering application value for the study of submarine sand waves.更多还原