Study On Extraction And Pattern Recognition Of Fashion Flat Sketches

At present, in the field of garment CAD system, the research on electronic Made To Measure(eMTM) mainly concentrate in the acquisition of the human body size, sample quickly generation and virtual trial system of garments. The fashion flat sketches mainly come from fashion designers. This thesis mades a systematic study of extraction and pattern recognition of fashion flat sketches using the computer vision technology, which is a functional extension of traditional garment CAD system. The main works and approach in the study can be outlined as following.(1) The method for clothing segmentation and its quality assessmentThe proposed approach adopts tow segmentation algorithms which can produce initial oversegmentation results, mean shift and the state-of-the-art segmentation algorithm, gpb-owt-ucm, combined with maximal-similarity based region merging(MSRM) method aiming to clothing segmentation. The fashion model image is transformed into a hierarchy of regions by using gpbowt-ucm algorithm. Meanwhile, the fashion model image is transformed into an initial oversegmentation result by using mean shift algorithm. Then the hierarchy or the initial oversegmentation result serves as a natural starting point for interactive segmentation based on MSRM, and then G-SSIM is used as the measure for the quality assessment of clothing segmentation.The hierarchical segmentation algorithm, gPb-owt-ucm, transforms the output of contour detector gPb into a hierarchical region tree, which can obtain the initial over-segmentation result by adjusting the boundary probability threshold. Mean shift is a procedure for locating the maxima of a density function given discrete data sampled from that function. It is useful for detecting the modes of this density. Unlike gPb-owt-ucm, mean shift can directly produce an initial oversegmentation result in the image segmentation process. The initial over-segmentation result produced by gPb-owt-ucm hierarchical segmentation algorithm is superior to that by mean shift in two aspects. First, fashion model images often contain complicated texture. gPb-owt-ucm algorithm has a certain advantage for this task in this satuation. Second, we can control the granularity of segmentation by thresholding the UCM without having to rerun the segmentation procedure.In recent years, region merging based image segmentation approaches attracted more research efforts. From an initial over-segmentation result, the basic operation of these approaches is to progressively merge similar neighboring regions into new regions according to some predefined merging criteria. The color histogram, which is produced by computing the parameters of RGB color space, is used to represent each region. The similarity measure ρ(R, Q) between each pair of adjacent regions, R and Q, is used as the merging criterion of maximal-similarity based region merging(MSRM) method.As an objective assessment of image quality, the structural similarity index CW-SSIM is a general purpose image similarity index which benefits from the fact that the relative phase patterns of complex wavelet coefficients can well preserve the structural information of local image features, and rigid translation of image structures will lead to constant phase shift. G-SSIM adopts the principle of CW-SSIM and slightly modifies it into a new one, which uses the complex Gabor filtering coefficients of an image instead of the steerable complex wavelet transform coefficients. The initial segmentation results are regional merged to the final segmentation results by MSRM method. We campare the two types of the final segmentation results produced by g Pb-owt-ucm and mean shift respectively with Ground Truth images extracted manully through Photoshop and give an objective assessment of the two types of the final segmentation quality.(2) Extraction of fashion flat sketchesThe extraction algorithm includes two parts: contour extraction and clothing internal details extraction.Contour extraction algorithm including initial contour extraction, branching contour smoothing and global contour smoothing processing is proposed. In the case of similar gray value of garment with printed patterns and background, contour extraction will produce texture noise. Introducing such concepts as contour error and branch point, a new initial contour exaction method is put forward and applied to garment images with printed patterns. Firstly, the initial contour is extracted based on the morphology, and then the initial contour is divided into several single-value branches. Finally, the texture noises in the branch are eliminated based on the contour error computation. The results indicate that the initial contour extraction algorithm based on the morphology can effectively remove image pixels within the garment images and the shadow and watermark outside. The contour error computation technology can effectively identify the texture noise data in the branch and do repairs for minimizing the effects of printed patterns on the garment contour. Meanwhile contrasting to other method,the proposed algorithm can realize the contour extraction with higher efficiency. Then, this paper realizes global contour smoothing by the following procedure:(1) Separating the curve;(2) Feature points extraction from separated curves;(3) Faire separated curves.Clothing internal details extraction algorithm is based on global contour smoothing for subsequent processing. we need symmetric point calculation. Many garment design draws are bilateral symmetric. Setting symmetric point can avoid dislocation when the left part is copied to the right or on the contrary. Finally, we extract style curves and find their intersections. We select sampling pixels Ai(i=1,...,m) from style curve in the original image through the mouse interactive operation, where m is the sampling number, then fit style curve by a cubic smoothing spline. Usually selecting pixels through mouse interactive operation cannot accurately locate two endpoints of the curve segment, ends of Ai extending for a length of l is required, namely, j=1-l,...,m+l. Then the curve B is discretized into Bj. Then searching Bj from one end to the other to find intersections, whose pixel value equals 1, the value of intersections is set to 2. Finally, the spare part of Bj on both sides should be removed.(3) Feature extraction and pattern recognition of fashion flat sketchesAn integrated approach to fashion flat sketches classification is proposed. Discrete wavelet and Fourier transform are employed to transform fashion flat sketches to a high-dimensional feature space. Then, LDA is adopted to map the high-dimensional features to low-dimensional feature space under the multi-class classification situation. At last, ELM is taken as the classifier.The proposed single scale of WFD, compared with other two kinds of the same length of the descriptor, has higher classification accuracy with or without dimension reduction. At the same time, it can be found in the experiment that the longer the length of the descriptor of the sample, the lower the classification accuracy. However, Samples with longer length of the descriptor after LDA dimension reduction has the higher classification accuracy. The same descriptor with the same length after dimention reduction, the classification accuracy of the LDA is higher than that of PCA. In order to avoid the problem of small sample, the length of the descriptor can’t take too long in dimention reduction by LDA.As observed from the experimental results, the traditional approaches of FD-SVM, MFDSVM and WFD-SVM achieve their best classification accuracy below 90 percent. As to the efficiency of feature extraction methods, we only need to compare WFD-LDA with FD-LDA because MFD-LDA easily causes small sample size problem. From the experimental results of classification accuracy comparison, WFD-LDA is more superior. The last experimental results show that the proposed approach of WFD-LDA-ELM is more accurate and faster than the approach of WFD-LDA-SVM.The proposed approach can be incorporated into the information system for large-scale garment design companies and lay the foundation for the garment style query in the next step. Moreover, this method can be applied to other classification problems, such as the fiber crosssectional shape classification, which are based on boundary description technique.(4) Lapel pattern recognition in fashion flat sketches based on lapel modelThe traditional feature extraction is highly dependent on clothing segmentation, which is difficult to achieve. The goal of the research of this section achieves the identification of clothing component in fashion flat sketches without clothing segmentation. Taking lapel as example, the purpose of this paper is to study the problem of recognizing the same clothing component in the flat sketches. Firstly, suppose that the four lines in a lapel are concurrent and symmetrical, a mathematical lapel model is built. Then, a novel method for some kinds of lapels recognition in fashion flat sketches based on lapel model is proposed.In the image preprocessing stage, the images need to be cropped for two times in order to remove the background and extract the region of interest, respectively. In the concurrent recognition stage, Hough transform, theta selection and curve fitting are applied to limit candidate lines. In the symmetrical recognition stage, k-means clustering algorithm is employed to partition lines to four clusters. The threshold values of the difference of corresponding weighted theta are set as lapel recognition criterion. Experiments demonstrate that the recognition accuracy of the proposed method is obtained at about 91.7%.(5) The integrated retrieval of flat sketches based on CBIRIn order to achieve the integrated retrieval of flat sketches, this paper takes advantage of the content-based image retrieval(CBIR) technology based on a database composed of 1229 flat sketches images of JPEG format. The single scaling wavelet Fourier descriptors, obtained form boundary data that were decomposed into second level using db10 wavelet with the length of the top 40, are used to describe shape features. Texture feature can be divided into global texture and collar, placket and hem local texture, using wavelet moment as global or local texture feature description. Shape and different texture features can be selected as an optional combination. Taking the optional features combination including different local texture features can find some flat sketches which cannot be retrieved only on shape and global texture features.