DMD-YOLO: a lightweight model for detecting surface defects on ceramic tiles

Abstract

Ceramic tiles are essential materials in architectural decoration, and their surface quality directly affects product aesthetics and market value. During manufacturing, various defects such as cracks and scratches may occur, characterized by diverse categories, large scale variations, and complex textures. However, existing detection models struggle to achieve a favorable balance between accuracy, efficiency, and lightweight design, often resulting in false positives and missed detections, particularly for small-scale defects under complex texture conditions. To address this, based on the YOLOv10 model, DMD-YOLO is proposed. The main improvements are as follows: (1) Introducing a reparameterizable DBB module and Mish activation function to enhance feature representation capabilities; (2) Employing the Monte Carlo Attention Mechanism and lightweight Dynamic Upsampling operator (DySample) to optimize the feature fusion process; (3) A small-object detection head (P2) and an ultra-lightweight detection head, Detect_dyhead, are designed to enhance the representation of small-scale defects; (4) A Shape-NWD loss function is designed to optimize bounding box regression by jointly considering shape and scale factors, thereby further enhancing the modeling capability for complex defects. Experimental results show that DMD-YOLO achieves an mAP50 of 79.6%, outperforming the baseline by 9.3%, with substantial improvements in small-sized defect detection. Furthermore, comparative experiments show that DMD-YOLO surpasses existing methods in terms of mAP50 while achieving a more favorable trade-off among model complexity, model size, and inference efficiency. In addition, evaluation on the NEU-DET dataset demonstrates that DMD-YOLO preserves strong detection performance on other industrial defect datasets, suggesting its potential generalization ability across multi-scale defect scenarios.