Abstract: Abstract: Aiming at the problems of insufficient adaptability due to reliance on fixed thresholds and manual rules of traditional machine vision methods in detecting defects on tile surfaces with complex textures, as well as the limited receptive field, high computational complexity, and poor performance in micro-defect detection of existing deep learning methods, this paper proposes a tile surface micro-defect detection method based on improved detr with enhanced matching for fast convergence. The method first constructs a frequency-aware backbone network by introducing wavelet down-sampling, wavelet convolution, and frequency domain attention mechanisms to preserve high-frequency details and suppress background interference. Secondly, a regional self-attention module is employed to replace the standard Transformer encoder, enhancing the modeling capability for local defect features while reducing computational complexity. Experimental results on the Ali Tianchi Tile Dataset show that the proposed method achieves an improvement of 8.7 and 1.7 percentage points in detection precision (mAP@0.75 and mAP@0.5, respectively) compared to the baseline model DEIM-s. In the detection of the most challenging micro-defects, the precision for dark dot defects improves by 16.8 and 2.7 percentage points in mAP@0.75 and mAP@0.5, respectively. Compared with mainstream models such as YOLOv8 (You Only Look Once) and RT-DETR (Real-Time Detection Transformer), the improved method proposed in this paper achieves the best results in terms of accuracy. This method effectively addresses the difficulties of feature loss and inaccurate localization of micro-defects under complex texture backgrounds, providing a new technical solution for high-precision and high-efficiency visual inspection in industrial tile quality control.

Key words: Keywords: deep learning, tile surface detection, small object detection, frequency awareness, detr with improved matching for fast convergence (DEIM)

摘要： 摘 要: 针对传统机器视觉方法在复杂纹理瓷砖表面缺陷检测中依赖固定阈值与手工规则适应性不足，以及现有深度学习方法存在感受野有限、计算复杂度高、对微小缺陷检测效果不佳的问题，提出一种基于改进DEIM (Detr with improved matching for fast convergence)的瓷砖表面微小缺陷检测方法。该方法首先构建频率感知主干网络，引入小波下采样、小波卷积与频域注意力机制，以保留高频细节并抑制背景干扰；其次，采用区域自注意力模块替代标准Transformer编码器，在降低计算复杂度的同时增强对局部缺陷特征的建模能力；最后，设计双阶段扩散聚集结构，通过双向跨尺度融合与多感受野特征提取，提升对多尺度缺陷的表征一致性。在阿里天池瓷砖数据集上的实验结果表明，所提出方法的检测精度mAP (mean Average Precision)@0.75和mAP@0.5相较于基线模型DEIM-s分别提升8.7和1.7个百分点；在对最具挑战性的微小缺陷检测上，对深色点缺陷的检测精度mAP@0.75和mAP@0.5分别提升16.8和2.7个百分点；与YOLOv8 (You Only Look Once)、RT-DETR(Real-Time Detection Transformer)等主流模型相比，本文改进方法在精度上均取得最优结果。该方法有效解决了复杂纹理背景下微小缺陷特征易丢失、定位不准的难题，为高精度、高效率的工业瓷砖视觉质检提供了新的技术方案。

关键词: 深度学习, 瓷砖表面缺陷检测, 小目标检测, 频率感知, DEIM