Abstract: To address the instability and subjectivity inherent in traditional tile color difference detection methods, a method integrating texture and color features was proposed for quantitative assessment and grading of chromatic aberration in tiles. A large-scale dataset named TILE-TCR (TILE Texture and Color Recognition), annotated with both texture and color labels, was constructed to enhance the model’s ability to represent these features. Additionally, the TILE-CAG (TILE Chromatic Aberration Grade) dataset was established to support the chromatic aberration classification task. Based on these datasets, the Fractal Encoding Network (FENet) was improved by incorporating Spatial Pyramid Pooling (SPP) and Squeeze-and-Excitation (SE) modules, enabling multi-task feature extraction and focusing on critical regions. A clustering algorithm was employed to adaptively determine the thresholds for chromatic aberration grading, allowing for objective quantification. Experimental results showed that the improved model achieved an accuracy oof 92.82% in the tile texture classification task, representing a 1.84 percentage point improvement over the baseline model. In the chromatic aberration grading task, all major evaluation metrics exceeded 90%. Furthermore, a simulated production line was built for industrial deployment and real-time performance testing. On commonly used tile specifications, the average detection time was under 3 seconds, meeting the real-time requirements for online inspection on industrial conveyor belts.

Key words: tile color difference detection, deep learning, multi-task learning, color difference grading, intelligent detection system

摘要： 针对瓷砖色差检测中传统方法依赖主观目测、结果不稳定的问题，提出一种融合纹理与颜色特征的瓷砖色差量化与分级方法。首先，构建了包含纹理与颜色双标签的大规模瓷砖数据集TILE-TCR（TILE Texture and Color Recognition），以提升模型对纹理与颜色特征的表征能力；同时构建色差分级数据集TILE-CAG（TILE Chromatic Aberration Grade dataset），用于支持色差分类任务。在此基础上，改进分形编码网络（FENet）的网络结构，引入空间金字塔池化（SPP）与SE（Squeeze-and-Excitation），实现多任务特征提取与关键区域聚焦。通过聚类算法自适应确定色差分级阈值，实现色差分级的客观量化。实验结果表明，改进模型在瓷砖纹理分类任务中准确率达到92.82%，较基准模型提升1.84个百分点；在色差分级任务中多个评价指标均超过90%。此外，搭建了模拟生产流水线，完成模型的工业部署与实时性测试，在常见规格瓷砖上平均检测时间低于3 s，满足工业传送带在线检测的实时性要求。

关键词: 瓷砖色差检测, 深度学习, 多任务学习, 色差分级, 智能检测系统