Abstract:

Objective Fabric surface defects influence the textile output, quality, price, and other factors directly, and it is hence necessary to devise a method for detecting fabric surface defects quickly and accurately in automatic production lines. This research aims to establish a statistical algorithm to achieve rapid detection of fabric surface defects.

Method Partial defects on the fabric surface could destroy its periodic geometric and statistical characteristics. Based on this feature, a detection method combining with an improved RANSac, named multi-metric-multi-model image voting (MMIV), was proposed. The input image was firstly divided into sub-images of the same size, and the output value matrix of the sub-image multi-dimensional metric was calculated. They were different from the multi-dimensional measurement standard values of the flawless background calculated by the improved Zero-Slope-RANSac method, and the basic scores of each sub-image were obtained by voting. Then the comprehensive scores obtained under the 4 counting models(square of standard mean, Borda, Copeland, Maximin) were sorted, finally, the defect sub-image represented by the outer point was obtained at the output end according to the sequence and offset.

Result The tested subjects were the self-sampled fabric defect dataset. When the RANSac method parameter was set to 3 and threshold set to 2, the confidence was greater than 0.25 and the prediction accuracy of single-measure-single-model reached 89.3% on average. The prediction accuracy reached 95.6% when the gray mean measure and Borda ranking model were selected, which was the highest, while the square of standard mean model (SSM model) had the lowest accuracy. Accuracy under 4-measures-3-models showed the prediction of 2 565 grey fabric images with non-latticed texture, and that of 3 708 grey fabric images with latticed texture background. The confidences of the both tables were greater than 0.35, and the prediction accuracy of each model was compared with the values of RANSac parameter set from 1 to 3, and the threshold set from 1 to 4. The prediction accuracy of Borda, Copeland, and Maximin points of the last three counting models was better than that of the SSM method. The average prediction accuracy of the combined multi-metric-multi-model image voting reached 92.7%, demonstrating a significant detection effect. By means of comparing non-lattice and lattice, it can be seen that the optimization of the multi-metric-multi-model image voting strategy was not applicable to lattice texture for the time being. Under the condition that high detection accuracy can be guaranteed, the detection speed of ZS-RANSac with 200 iterations was more than 5 times that of 1 000 iterations, meanwhile, the detection time reached only 0.466 s, satisfying the real-time performance of pipeline work. Among the four prediction models, the SSM model was 10 times faster than the other three models, and the average time of the other three models were relatively close, reaching the fastest 0.135 s of the Copeland model. Considering accuracy and real-time performance, the Borda counting model demobnstrated the best results.

Conclusion It can be seen from the experimental results that, the proposed algorithm can detect defects on the fabric surface for periodic texture images quickly and accurately, and a new dichotomy labeled dataset for periodic texture grey fabric was created. The algorithm does not require a large amount of preliminary data for training, can overcome the problem of the lack of public datasets in the industry to a certain extent, and is suitable for real-time defect detection of solid color and striped background fabric. This technology is able to reduce the labor cost of the factory in the industrial entity under certain circumstances, and provides an idea to apply the software statistical prediction method to the research of image algorithms. Future work would focus on metric screening and model optimization in multi-metric-multi-model image voting(MMIV),as well as adaptive optimization of RANSac involved parameters, to further improve detection accuracy and average detection speed.

Key words: object detection, periodic texture, fabric surface defect detection, zero-slope-RANSac, multi-metric-multi-model image voting