Accurate welding defect detection plays the key role to the welding quality control, which is also the basis of maintenance decision. However, the welding images from X-ray imaging system are always against poor texture and low contrast phenomenon, which causes weak feature representation ability. Meanwhile, the welding defects always occupy a small proportion of image pixels compared with backgrounds to bring the class imbalance issue. These complex factors bring a certain challenge to accurate welding defect detection. Recently, due to strong context feature representation ability, deep convolutional neural network (DCNN) has acquired a remarkable performance on defect segmentation. Nevertheless, high-precision defect segmentation based on DCNNs is still a challenging task due to insufficient processing of local contextual feature maps, limited receptive field, etc. To address these issues, considering the encoder–decoder framework, an effective welding defect segmentation network is proposed for end-to-end defect location from X-ray images. Specifically, an effective backbone with a bidirectional convolutional long short-term memory (BiConvLSTM) block is built to learn the global, long-range contexts and improve the network’s propagation ability of subtle context features. Meanwhile, to address the insufficient processing issue of local contextual feature maps, to imitate human visual attention, a global attention block is proposed for local feature enhancement to make the segmentation network emphasize the defective areas. In addition, aimed at the limited receptive field, a feature enhancement block is proposed for multi-scale feature representation and fusion. Experiments on public dataset of X-ray welding defects show that the proposed defect segmentation network could acquire a promising segmentation performance compared with other related segmentation models.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号