基于DenseNet网络的焊缝缺陷检测识别

Weld defect detection and recognition based on DenseNet network

针对X射线焊缝的缺陷分类识别难度较高且传统算法复杂、低效的问题,引入了基于深度学习的密集连接卷积网络(DenseNet)算法,并对数据进行了动态增强。DenseNet网络算法脱离了机器学习算法中需要加深网络层数和加宽网络结构来提升性能的定式思维。通过特征重用和旁路设置,从而实现对焊缝缺陷的检测识别。在相同数据集和训练步数下,同最小二乘支持向量机(LS-SVM)与卷积神经网络LeNet算法对比,DenseNet网络提高了模型泛化能力和识别准确率,对焊缝缺陷识别准确率可达98.969%。

In order to solve the problem of high difficulty of defect classification recognition of X-ray weld seam,low efficient and complicated that traditional algorithms is dense convolutional network(DenseNet)algorithm based on deep learning is introduced,and the data are dynamically enhanced.Dense Net network algorithm breaks away from the stereotyped thinking that needs to deepen network layers and widen network structure to improve performance in machine learning algorithm.Through feature reuse and bypass setting,the detection and recognition of weld defects can be realized.Under the same data set and training steps.compared with LS-SVM and LeNet algorithm of convolution neural network,DenseNet network improves the generalization ability and recognition accuracy of the model,and the recognition accuracy of weld defects can reach 98.969%.