HIP-FCA-Net: an intelligent detection method for pixel-level fine cracks on the surface of castings integrating image adaptive enhancement function

Abstract Fine cracks are prevalent surface defects in cast components. Their effective detection is critical for ensuring metal part quality. However, complex industrial environments pose significant challenges for efficient automated pixel-level detection of such defects. To address this issue, we propose an intelligent detection method based on machine vision, termed hybrid image preprocessing (HIP)-fine crack attention (FCA)-Net. The proposed method comprises two primary components: a HIP module and a FCA-integrated detection network (FCA-Net). HIP employs multiple image processing units with adaptively adjustable parameters. It enhances the contrast between cracks and the background while suppressing noise and artifacts, thereby providing more discriminative inputs. FCA-Net employs U-Net as the backbone and incorporates an attention mechanism specifically designed for fine cracks. This mechanism strengthens pixel-level feature extraction and representation along elongated crack edges. Furthermore, Dice loss is introduced during training to mitigate severe class imbalance and improve segmentation accuracy. Experimental results on a self-constructed casting surface crack dataset demonstrate that the proposed method achieves a pixel accuracy (PA) of 92.33%, an intersection over Union of 75.31%, a Dice coefficient (Dice) of 85.92%, and a mean average precision of 85.52%. Compared with existing methods, HIP-FCA-Net exhibits significant improvements across multiple key performance metrics. It offers an effective and practically viable solution for online intelligent detection of casting surface defects.