【摘要】 邮轮薄板焊缝的熔深和熔宽相对较小，母材与焊缝区域差异性小，焊缝表面缺陷较难判别。为准确地定位焊缝位置，提出一种将注意力机制的坐标注意力模块(Coordinate Attention Module, CAM)融入密集链接卷积网络(Densely Connected Convolutional Networks, DenseNet)的邮轮薄板焊缝缺陷识别算法，建立CAM-DenseNet模型。将网络中的激活函数ReLU替换为更具有稳定性的ReLU6,并利用贝叶斯优化算法对CAM-DenseNet模型的超参数组合进行优化和选取。在焊接车间利用相机采集邮轮薄板焊缝三原色(Red Green Blue, RGB)图片，自建立邮轮薄板焊缝缺陷数据集，并按焊缝缺陷类型将数据集分为凹陷、气孔、毛刺、表面裂纹和无缺陷等5类。试验结果表明，CAM-DonseNet模型对邮轮薄板焊缝缺陷识别具有优异表现。更多还原

【Abstract】 The depth of fusion and width of fusion of weld of cruise ship sheet are relatively small, the difference between base material and weld area is small, and the weld surface defects are difficult to identify. In order to locate the weld position more accurately, an identification algorithm for weld defect of cruise ship sheet with intergrating the Coordinate Attention Module(CAM) of attention mechanism into Densely Connected Convolutional Networks(DenseNet) is proposed, and a CAM-DenseNet model is established. The activation function ReLU in the network is replaced by ReLU6 with more stability, and the hyperparameter combination of CAM-DenseNet model is optimized and selected with Bayesian optimization algorithm. The Red Green Blue(RGB) pictures of weld of cruise ship sheet are collected with the camera in the welding workshop, the data sets of weld defects of cruise ship sheet are self-established, and the data sets are divided into 5 types according to weld defect types, such as concave, porosity, burr, surface crack, and no defect. The test results show that the CAM-DenseNet model is of the excellent performance in weld defect identification of cruise ship sheet.更多还原