A robust model for surface multi-class defect detection using interactive network architectures

Current methods for identifying surface defects often struggle to balance sensitivity to subtle defects, modelling of multi-scale context, and computational efficiency. This limits their application in real-time inspection systems. In this paper, we propose a robust defect detection model based on YOLOv8 which introduces two novel components to enhance accuracy and adaptability. The first is the Attention-based Local Integration Block (ALIB). This incorporates an Enhanced Perception Block (EPB) that uses dynamic spatial-channel modulation to highlight subtle defect features. The EPB strengthens defect responses and suppresses background interference. The second component is Shared Weight Dilated Fusion (SWDF). This component efficiently captures and merges multi-scale features, enlarging receptive fields with minimal parameters while preserving spatial details. Evaluations on four benchmark datasets and one mixed dataset demonstrate that our model improves mean average precision (mAP) by 4.8%, 5.2%, 2.1% and 3.1%, respectively, and achieves an increase in mAP@0.5:0.95 of 5.3%, 0.6%, 1.4% and 3%. With a computational cost of only 7.8 GFLOPs and a small memory footprint. These results demonstrate that our detector effectively balances accuracy and efficiency. It can easily be adapted for use in embedded vision systems in automated production environments, real-time inspection where high precision and speed are required.