Balanced multi-scale target score network for ceramic tile surface defect detection

Abstract

Ceramic tiles, as a prevalent building material, exhibit a wide variety of types and high demand. Traditional manual inspection methods relying on human visual observation suffer from low efficiency and unreliable accuracy. Current automated detection methods mostly rely on traditional image processing techniques for feature extraction, followed by machine learning-based classification. However, faced with the diversity of tile types and defect categories, fine-tuning and deployment processes require significant human and material resources, while detection efficiency remains limited. In this study, we first construct a high-resolution dataset for studying surface defects in ceramic tiles (CT surface defects dataset), encompassing multiple batches and various patterns of tiles. Subsequently, data analysis is conducted to address the scale and quantity differences in defect distribution. We propose an improved approach by introducing a content-aware feature recombination method and a dynamic attention mechanism to enhance the classical single-stage object detection algorithm YOLOv5. These enhancements aim to reduce information loss in features and enhance the expression of multi-scale features. Furthermore, we design a loss function that mitigates score differences for multi-scale defects. The proposed approach mitigates the discrepancy in contribution among different scale targets caused by imbalanced quantities. It effectively prevents the model from excessively favoring a specific scale target during the learning process. Experimental results demonstrate the superior accuracy and efficiency of our detection method.Compared to the baseline network YOLOv5, our approach achieved improvements of 4.9% in AP (Average Precision), 6% in APs (small-scale objects), and 8% in APl (large-scale objects). Furthermore, we achieved a 3.9% improvement in detecting white point defects, which are most affected by small-scale objects, and a 4.1% improvement in detecting discolored spot defect, which are most affected by class imbalance.