Multiscale welding defect detection method based on image adaptive enhancement

The automatic detection of welding internal defects using radiographic images is an important technique for improving the efficiency and consistency of weld fault diagnosis. However, accurate defect detection is challenging due to the low contrast of radiographic images and the large difference in the sizes of different welding defects. In existing methods, the ray image enhancement and defect detection processes are isolated, and the enhancements that are beneficial to defect detection need to be obtained by manual parameter adjustment, which cannot adapt to large-scale detection tasks. Moreover, the adjustment strategy of the methods to the input image is not conducive to detecting multiscale welding defects. Therefore, this paper proposes a multiscale welding defect detection method based on image adaptive enhancement to address these problems. The method comprises two modules: image adaptive adjustment (IAA) and defect detection based on global and local semantic fusion (DD-GLF). In the IAA module, the parameter prediction network is trained to adaptively predict the parameters of the differentiable image processing function to improve the detection accuracy, and in the DD-GLF module, a defect detection model that accepts global and local window images of welds as inputs is designed to detect multiscale welding defects. Experiments on actual inspection data show that the proposed method achieves enhancement results that are consistent with those of human experts and performs well for dense and large defects.