Casting defect region segmentation method based on dual-channel encoding–fusion decoding network

Segmenting casting defect regions is vital for assessing defect levels in casting products. In complex backgrounds with multi-scale defects and regions of fuzzy and weak texture, existing methods often fail to capture detailed features of the defect, leading to incomplete segmentation. This study developed a casting defect region segmentation method based on a dual–channel encoding–fusion decoding (DCE–FD) network. Initially, an encoding network module based on a deep and shallow dual–channel structure (ENM–DSDCS) was established to extract macroscopic and multi-scale detail features using deep and shallow structure networks, respectively. This approach ensures comprehensive feature extraction from multi-scale fuzzy and weak texture defect areas. Further, a decoding network module based on attention-based bidirectional guidance fusion (DNM–ABGF) was developed. This module guides the semantic and multi-scale detail branch features to achieve complementary information fusion at each scale level during the decoding stage, thereby preserving fuzzy boundaries and other details in the fusion process and enhancing the accuracy and integrity of segmentation. Experimental results demonstrate that the mean intersection over union (mIOU) and Dice coefficients for defect region segmentation in radiographic images of castings were 92% and 75.71%, respectively. These metrics surpass those of eight advanced segmentation methods in terms of accuracy and consistency. The proposed method significantly improves sensitivity to feature detail and accurately segments potential fuzzy regions of multi-scale defects, offering promising advancements in the nondestructive testing of casting products.