Abstract Surface defect segmentation (SDS) is challenging in automated inspections due to the high variability of defects amid complex metallic textures. The defects show significant variations in size, shape, contrast, and spatial distribution. Existing segmentation methods struggle to jointly model fine boundary cues, omni-directional long-range dependencies, and multi-scale context in a single framework. To address these issues, we propose Edge-Guided Omni-Directional Attention Network (EGONet), a novel hierarchical architecture for metallic SDS. EGONet offers rich, multi-level intermediate features that capture both fine-grained details and broader semantic information. It also enhances multi-scale contextual understanding by integrating a Dense Atrous Spatial Pyramid Pooling (DASPP) module with four parallel dilated-convolution branches and a learnable residual blend for maintaining spatial resolution. A Sobel-guided Edge Attention Module (EAM) highlights high-frequency boundary cues at the shallowest feature level via auxiliary edge supervision. The proposed Omni-Directional Attention (ODA) decomposes spatial attention into four axes to model long-range dependencies across all orientations via a two-stage refinement. A Quad-Statistical Spatial Attention Module (SAM) leverages quadruple pooling to deliver finer spatial sensitivity than traditional dual-pooling methods. Moreover, the Efficient Channel Attention (ECA) module recalibrates channel responses to suppress background interference. Lastly, a Bilateral Feature Integration Block (BFIB) combines the dual attention pathways via complementary asymmetric gating and residual stabilization. Extensive ablation studies validate each component’s contribution toward improving overall detection performance. Experiments on two benchmark datasets demonstrate that EGONet achieves competitive performance against state-of-the-art methods, with leading results across multiple defect categories on MT-Defect and strong generalization on SD900. Subjects Introduction Surface defect detection (SDD) is critical for maintaining product quality and operational efficiency in modern manufacturing. Conventional inspection methods suffer from fundamental limitations in precision and adaptability to complex patterns. Human-based inspection is inherently slow and prone to inconsistent judgments1. Ultrasonic methods are constrained by material acoustic properties and require