PercepWeldNet: A Novel Perception-Guided Deep Model for Welding Defect Detection

In this study, we introduce PercepWeldNet, a deep architecture designed for visual perception in welding defect detection, addressing the challenge of identifying semantically significant features within complex welding scenes. Our approach integrates multi-channel perceptual visual features to effectively represent welding images with intricate spatial configurations. The core contribution is a multi-layer hierarchical model, LPL, which simulates human gaze behavior by utilizing the BING objectness metric. This enables efficient identification of semantically and visually important patches across various scales and contexts. Through locality-preserved learning (LPL), our model generates gaze shift paths (GSP) that preserve local coherence and focus on regions that align with human visual attention. These GSP-based deep features are incorporated into a Gaussian mixture model for welding defect classification. Extensive empirical validation on our welding defect detection dataset and several generic datasets demonstrates the effectiveness of our approach, with significant advantages in defect categorization accuracy. The proposed method achieves superior results in welding defect detection, as well as in scenery retargeting tasks.