Abstract:

Objective Textile fabric defects are generally caused by raw materials (warp or weft yarns), mechanical failures and human factors in the production process, and they seriously impair the quality of final products. At present, most of the defect inspection is conducted by human inspectors, resulting in low efficiency and high laboring cost. Therefore, it is of great significance to apply fast and reliable image processing and machine vision techniques to perform automated defect detection instead of human.

Method Sparse dictionary learning method has excellent adaptability in representing complex fabrics textures. However, the learning and solving of sparse dictionary take a long time, making it hard to meet the real time requirements in the industrial scenarios. This work proposed a novel dictionary grouping strategy to speed up the sparse coding process in detection stage while guaranteeing the detection accuracy. Firstly, the sliding patch scheme was adopted to learn dictionary from normal fabric samples. Secondly, the learned dictionary was optimized by dividing into groups, and such strategy is to select several optimal dictionary groups with respect to the degree of approximation. Next, the optimized dictionary groups were used to reconstruct a test image to obtain its residual image. As the final step, the reconstructed errors were applied to identify defect areas from normal ones.

Results To compare the computation time of different algorithms, only the total running time in detection stage was calculated, not including the dictionary learning or dictionary grouping. The experimental results showed that the sparse dictionary algorithm took a longest running time among them, the proposed algorithm took the second longest time, and the unconstrained dictionary used the shortest time(Tabl. 1). The reason that the proposed algorithm was able to reduce most of the time is that the entire algorithm process advances the process of finding the optimal dictionary atoms for the sparse dictionary (sparse coding) and limits the number of dictionary atoms and the reconstruction error, thus significantly reducing the computational effort. The proposed algorithm also achieved a high accuracy rate, slightly lower than the sparse dictionary. The possible reason for it may be that the sparse dictionary do not use all dictionary elements patches each time, but selecting the least number of elements for patch approximation, helping ignore details such as defective areas. The three dictionary learning algorithms have relatively good detection results for various types of defects in plain gray fabrics, and the detection rates are above 90% (Tab. 2). This proposed method has excellent and stable detection for warp defects such as broken warp and knots, and poor detection for weft defects only for few images, which may be caused by their low contrast or minor anomalous defective areas. In summary, the detection accuracy of the proposed method is comparable to that of the unconstrained dictionary and the sparse dictionary in terms of detection accuracy.

Conclusion In this work, the grouped sparse dictionary method has been proposed to address the real time flaw inspection problem on textile fabric. Aiming at the time consumption of spare coding process, the large amount of spare dictionary are grouped into several groups with small dictionary atoms inside in advance. By converting the time-consuming sparse coding into a least square problem, the proposed method has been proved to be capable of reducing computation time in inspection phase significantly. The experimental results show that this method can combine the advantages of learning time of unconstrained dictionaries and high accuracy of sparse dictionaries, while ensuring real-time and low false detection rate, and has strong adaptability to different types of defects, especially for warp defects with high accuracy and stability.

Key words: sparse representation, dictionary optimization, fabric defect, real-time detection, image optimization