Abstract: Four novel fractal features are extracted based on the analysis of the problem existing in estimation of fractal dimension by box-counting method as well as on the analysis of the inherent characteristics of woven fabrics.Feature vector composed of these four features together with Euclidean measure are then applied to the detection of defects in seven fabrics of different textures background.Testing results indicate that，using the proposed algorithm，actual false alarm probability and missing rate can be controlled simultaneously below 10% under an appropriate range of threshold，proving with convince the effectiveness of the fractal features extracted in this study.

Key words: fabric defects detection, novelty detection, fractal dimension, box-counting method

摘要： 深入分析了计盒法在估算分形维数上存在的问题并结合织物纹理固有的特点，提取了4个新的分形参数，利用由它们组成的特征向量并使用欧式距离检测器对7种具有不同纹理背景的织物进行了瑕疵检测。测试结果表明，在一定的阈值范围内，实际的误警率和漏检率可以同时控制在10％以内，有力地证实了所提取的分形特征参数的有效性。

关键词: 织疵检测, 新奇检测, 分形维数, 计盒法