基于改进注意力模块的船舶涂装缺陷检测方法

Ship painting defect detection method based on improved attention module

针对人工检测船舶缺陷效率低、传统检测网络准确率差的问题,提出一种基于改进注意力模块(improved convolutional block attention module, ICBAM)的船舶涂装缺陷检测方法.首先,YOLOv4在路径聚合网络中将深度可分离卷积代替常规卷积形成IYOLOv4,减少模型计算量;其次,将ICBAM融入IYOLOv4的路径聚合网络Route层后形成ICBAM-IYOLOv4,ICBAM在通道上构建多频率通道改善全局平均池化,利用一维卷积代替全连接层聚合相邻通道间的信息,减少模型参数;然后,在空间上融合Inception v3思想和特征分层思想改善空洞卷积;最后,在船舶涂装缺陷样本数据增强的基础上,对ICBAM-IYOLOv4进行测试.实验结果表明:ICBAM-IYOLOv4相比其他算法,其损失值更低、收敛更快;平均精度均值(mean average precision, MAP)在训练集和测试集上分别提高了1.89%和1.91%.

To address the problems of low efficiency of manual detection of ship defects and poor accuracy of traditional detection networks,a ship painting defect detection method based on Improved Convolutional Block Attention Module(ICBAM)is proposed.Firstly,YOLOv4 replaces the conventional convolution with depthwise separable convolution in the path aggregation network to form IYOLOv4,which reduces the model computation;secondly,ICBAM is incorporated into the Route layer of the path aggregation network of IYOLOv4 to form ICBAM-IYOLOv4,ICBAM builds multi-frequency channels on the channels to improve global average pooling,and one-dimensional convolution is used instead of the fully connected layer to aggregate information between adjacent channels and reduce model parameters;then,the idea of Inception v3 and the idea of feature layering were spatially fused to improve the dilated convolution;finally,ICBAM-IYOLOv4 is tested on the basis of enhanced data from ship painting defects samples.The experimental results show that ICBAM-IYOLOv4 has lower loss values and faster convergence than other algorithms,and the mean average precision(MAP,Mean Average Precision)is improved by 1.89%and 1.91%on the training and test sets,respectively.