服装图像自动标注方法研究

发布时间：2018-07-16 14:34

【摘要】：随着信息时代的到来,时尚品牌的服装网站服装图像数量剧增,出现了许多款式多样、没有分类明确的服装图像,导致服装工作者能够挑选出适用的、满足设计需求、给自己创作带来灵感的设计花费的时间和精力越来越多。服装相关工作者如何从海量的服装图像库中快速准确得检索到自己满意的图像是目前的一个热点研究领域。由于服装图像底层视觉特征和高层语义概念之间普遍存在理解偏差,人工标注服装图像已经难以满足用户需求,为了更好的实现服装图像检索技术,服装图像自动语义标注越来越占据着举足轻重的作用。服装图像自动标注技术中,采用训练好的分类器模型为底层视觉特征和高层语义概念建立联系,待标注服装图像则可根据建立好的联系模型实现服装图像的自动标注。本文围绕分类器模型技术,对服装图像自动标注及其相关技术进行了研究,主要包括颜色特征和纹理特征两方面的视觉特征的提取和算法描述,将提取的视觉特征归并融合为一个综合特征向量作为实验数据。然后通过研究支持向量机的原理和技术,并且在大量实验的基础上,提出了一种基于多核SVM服装图像自动标注方法,并结合交叉验证法调优参数,实现了一个基于SVM的服装图像分类和标注系统,此系统明显提高了服装图像自动标注的准确率。
[Abstract]:With the advent of the information age, the number of clothing images on fashion brands' clothing website has increased dramatically, and many fashion images have appeared in various styles, without a clear classification of clothing images, leading to the clothing workers to be able to select suitable clothing images to meet the design needs. More and more time and energy is spent on the design that inspires you. How to retrieve their satisfied images quickly and accurately from a large amount of clothing image database is a hot research field. Because there is a general understanding deviation between the low-level visual features of clothing images and high-level semantic concepts, manual tagging of clothing images has been difficult to meet the needs of users, in order to better achieve clothing image retrieval technology. Automatic semantic tagging of clothing images is playing a more and more important role. In the automatic tagging technology of clothing image, the trained classifier model is used to establish the connection between the underlying visual feature and the high-level semantic concept, and the garment image to be tagged can be automatically tagged according to the established contact model. Based on classifier model technology, this paper studies the automatic tagging of clothing image and its related techniques, mainly including the extraction of visual features and the description of algorithm from two aspects: color feature and texture feature. The extracted visual features are merged into a synthetic feature vector as experimental data. Then, by studying the principle and technology of support vector machine, and on the basis of a lot of experiments, this paper proposes an automatic tagging method based on multi-core SVM, and adjusts the parameters with cross-validation method. A clothing image classification and tagging system based on SVM is implemented, which improves the accuracy of automatic garment image tagging.
【学位授予单位】：北京服装学院
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TS941.2

【相似文献】