DLE-UNet: Cross-Group Enhanced Channel Attention and Dual-Path Collaborative Encoding for Defect Detection in Aluminum Castings

Abstract

Segmenting defects in X-ray digital radiography (DR) images of aluminum castings presents significant challenges due to the extremely low pixel coverage of minute defects, blurred boundaries, and weak contrast. The core issues lie in the fundamental trade-off between expanding the receptive field and preserving fine spatial details, as well as the background noise introduced by indiscriminate feature propagation during decoder fusion. To address these challenges, we propose a novel and improved U-Net architecture named DLE-UNet. First, we design a decoupled collaborative dual-path encoder. By structurally separating the contextual semantic stream from the spatial detail stream, it achieves synergistic optimization of large receptive fields and high-resolution details, thereby mitigating the loss of minute defect features in deep layers. Second, we introduce a cascaded boundary refinement decoder. In its initial stage, a Large-kernel Group Attention Gating module performs coarse feature screening to suppress background noise. Subsequently, a content-guided upsampling module conducts sub-pixel-level refinement of boundary features. Experiments on an aluminum casting defect dataset demonstrate that our method outperforms existing approaches on key metrics such as the F1-score (81.3%) and Intersection over Union (IoU, 68.4%). Notably, it maintains a superior recall rate while achieving high precision, validating its advanced capability and robustness in performing high-integrity, high-precision segmentation of minute defects in complex industrial scenarios.