关键词: 提花针织物, PM模型, 扩散方程, 梯度阈值, 改进局部二值算法, 去邻域归一化, 多特征融合

Abstract:

Objective Supervision and inspection are important parts in quality control of the knitted fabric production process. The defect detection by automation and machine vision technology can effectively improve the detection efficiency. Jacquard knitted fabrics have prominent yarn edges, obvious loop characteristics and patterns, which have a strong interference to the defect detection process. Therefore, it is necessary to design an effective and accurate pretreatment and defect detection method for the complex texture of jacquard knitted fabrics.

Method Improved PM (perona-malik)model was adopted to suppress the strong texture edge of jacquard knitted fabric. Firstly, the image was divided into texture and defect region by gradient difference before selecting the corresponding diffusion equation. The gradient threshold was determined according to the probability subset calculated by the gradient matrix and the relevant criteria to achieve regional diffusion. According to the texture distribution characteristics, the improved local binary pattern(LBP), entropy and correlation were extracted, and then the defect regions were further highlighted by neighborhood normalization and multi-feature fusion. Finally, the defect morphology was located and segmented by region growth method.

Results The effectiveness of the preprocessing method and the defect detection method for texture suppression was experimental investigated and analyzed, and defect information extraction of jacquard knitted fabric was demonstrated exhaustively. In addition, several different preprocessing algorithms and defect detection algorithms were compared and demonstrated. Comparison of defect image before and after preprocessing showed that the gray fluctuation amplitude of the image after preprocessing was smaller and the texture distribution was more concentrated. It was seen from the pretreatment experimental images and from the comparison effect with other preprocessing algorithms that the abrupt change of the texture edge area was still obvious, and the yarn spacing area of the fabric was still obvious in the visual effect. However, the proposed preprocessing algorithm effectively suppressed the strong texture edges and yarn spacing of patterns, and the intensity of texture edge filtering was greater. At the same time, in order to prove the robustness of the preprocessing effect more intuitively, the background suppression factor BSF, structure similarity SSIM and signal-to-noise ratio gain ISNR were used for index evaluation (Tab. 1). From characteristic reconstruction diagrams and normalized characteristic diagrams, the texture features selected in this paper effectively described the gray difference and gray distribution difference between defects and textures. In addition, the normalization of the row mean value without neighborhood effectively weakened the eigenvalue of the texture region and increased the difference between the texture and the defect. The effect of different defect detection algorithms on the defect image of jacquard knitted fabric was showed that, the algorithm in docoment [3] may misjudge the fabrics with single pattern or similar patterns, while the fabrics with multiple patterns may be interfered by complex patterns, resulting in false detection. The algorithm adopted in docoment[4] was not able to rule out that the interference of yarn coil spacing, which lead to missed inspection and false inspection. The algorithm in this paper achieved the localization of the defect area and extracts a relatively complete defect shape contour. The comparison results of the detection accuracy of each detection algorithm for 100 experimental images were obtained (Tab. 2). The false detection rate of the algorithm was 3.3% for normal fabric images and 98.6% for defective fabric images, further illustrating the effectiveness of the algorithm.

Conclusion Aiming at the complex texture of jacquard knitted fabric, this paper proposed a defect detection algorithm based on nonlinear diffusion and multi-feature fusion. The improved nonlinear diffusion model was used as the pretreatment means, and the single diffusion mode of the conventional PM model was improved to the regional diffusion by selecting the best diffusion equation and gradient threshold. At the same time, multi-feature extraction and fusion were used as detection means to further highlight the defect area by using without neighborhood normalization and weighted fusion methods, and finally the defect shape was segmented by using region growth method. The experimental results show that the improved PM model effectively weakens the complex texture of jacquard knitted fabric and eliminates the interference caused by texture. Feature extraction method and normalization method increase the difference between texture and defect, and further highlight the defect. Compared with other detection methods, the detection accuracy of this paper is higher for jacquard knitted fabric defect image, and the defect region segmentation is more perfect and accurate.

Key words: jacquard knitted fabric, PM model, diffusion equation, gradient threshold, improved local binary pattern, removed neighborhood normalization, multi-feature fusion