Smoothing Filter With Texture Suppression And Its Application In Fabric Pattern Recognition

With the rapid development of computer technology, textile industry is striding toward the direction of automation. Textile industry automation is not only highly automation of textile equipment, also including the related textile design, detection, classification, analysis and recognition, and many other aspects. With the improvement of image acquisition technology and the rapid development of photography technology, the acquisition of the textile surface information, such as fabric parameters analysis, organizational identification, the fabric quality inspection, fabric pattern recognition can be completed by the computer instead of the traditional manual retrieval. People can take advantage of image acquisition equipment such as high resolution scanner and camera to collect the fabric image, and then process them by computer technology. However, the fabric scan images are difficult to reflect the real information on the surface of the fabric. First of all, the fabric usually consists of yarns of different colors and materials. With the increase of different kinds of yarns, the fabric color is also becoming rich and including varied levels. But the gaps between yarns are usually darker and the regions of the same color yarns will form certain veins. Secondly, the fabric is not always flat. The fabric scan images can hardly reflect the real color of the yarns. Under the scanning light, there may be color transition between the center and edges of the same yarns as the fabric yarns is usually the cylindrical structure. Thirdly, there may be local color difference, even pseudo patterns due to irregular placement during the scanning. In addition, the fabric scan image usually is compressed by means of JPEG, which can result in that the image contour becoming blur. All these factors make it difficult to reflect the real content of the fabric image. It is also adverse to post-processing such as image structure extraction, fabric pattern design, fabric color analysis, image retrieval and pattern recognition. The preprocessing results often determine the accuracy and validity of these methods.In computer graphics, we normally use a smoothing filter for image pre-processing. Nowadays, there has been formed some classic smooth algorithms such as mean filter, median filter and Gaussian filter. These classic filters smooth the image, but also could cause unexpected edge fog and loss of important details. In image processing, edge is a very important feature which is usually used to describing and identifying target. Therefore, the ideal image smoothing should be not losing edges while smoothing the regions between such edges. Unfortunately, such an operator does not exist, because in general it is impossible to unambiguously determine which edges should be preserved, which is always the difficulty of image processing technique.With the development of digital image processing technology and the introduction of mathematical theory, there have been some edge-preserving filter, the main methods include anisotropic diffusion, bilateral filter, weighted least squares, L0 gradient minimization, total variation and relative total variation model. Those methods can produce good export in certain applications. However, as for the fabric scan images, they are failed to achieve satisfactory results.According to the characteristics of the fabric scan images, in this paper, we do some works in smoothness and identification for the fabric scan images. The main contents include the following aspects:Firstly, according to the characteristics of the fabric scan images, we proposed a filter for smoothing fabric images while suppressing textures and preserving fabric pattern structures, by means of a nonlinear combination of local color values and lightness gradients underlying CIE-Lab color space. The proposed filter combines three functions to change the weights of the pixels near to the center pixel in both domain and range: a function to smooth color intensity in spatial domain, a function to smooth color intensity and preserve color edges in range domain, and a function to suppress textures based on local lightness gradient. As the proposed method needs to smooth each pixel in fabric image, it’s slightly complex to calculate. Then, we accelerated the smoothness method by piecewise linearization.Secondly, this paper proposed a multi-feature fusion method of fabric pattern recognition. First the yarn textures are removed by a smoothing filter with texture suppression, and the filtered images are greyed simultaneously; then three types of features including edge direction histogram, SURF features of maximally stable extremal regions and features of gray-level co-occurrence matrixes are extracted, and feature database of sample images is established; finally, with sample image features library as the training objects, the classifier which fuses three types of pattern features together is constructed by Adaboost algorithm. The experimental results show that the fabric pattern recognition method of multi-feature fusion by Adaboost algorithm achieves preferable accuracy.This paper does some research on the pre-processing of fabric’s scanned images. This paper may do help to structural extraction of scanned images, fabric flower design, fabric color analysis, image retrieval and pattern recognition. This paper provides a new way for fabric image recognition through the study of the fabric pattern recognition.