摘要：Aiming at the problem of low generalization ability and accuracy under imbalanced train dataset of identifying weld defects of radiographic images, the improved resnet18 deep neural networks combining q-learning algorithm was proposed. in this method, five types of weld defects images were enhanced as the datasets including slag inclusion, gas pore, crack, lack of fusion and non-defects. Q-learning algorithm was proposed to identify the types of weld defects and a personalized reward function was constructed for the imbalanced ratio (IMR) of the training set. The improved ResNet18 network was designed to extract weld defects images features deeply. The agent used the weld defects images as the current state and five types of defects were used as an agent action. The agent performed recognition actions at each time step and return the value. Finally, the network learned the strategy of weld defects recognition under IMR train dataset. The experimental results showed that the improved ResNet18 network with Q-learning algorithm had good generalization ability, for the extremely imbalanced dataset (IMR=1/9). The recall rate still reached over 92% and the accuracy rate reached over 87% and the precision rate over 90%. ? 2022 IEEE.

ISSN号：9781665406482

卷、期、页：页328-332

发表日期：2022-01-01

期刊分区（SCI为中科院分区）：无

收录情况：EI（工程索引）

发表期刊名称：Proceedings - 11th Electrical Power, Electronics, Communications, Control, and Informatics Seminar, EECCIS 2022

通讯作者：高博轩,苗兴园

第一作者：赵弘

论文类型：会议论文

论文概要：高博轩,赵弘,苗兴园，Recognition of Radiographic Weld Defects based on Combining ResNet18 and Q-learning for Imbalanced Train Dataset，Proceedings - 11th Electrical Power, Electronics, Communications, Control, and Informatics Seminar, EECCIS 2022，2022，页328-332

论文题目：Recognition of Radiographic Weld Defects based on Combining ResNet18 and Q-learning for Imbalanced Train Dataset