基于超像素分割的服饰提取算法研究与实现

发布时间：2018-02-09 20:40

本文关键词： 超像素分割区域比较服饰提取服饰属性图像分割　出处：《西南交通大学》2016年硕士论文　论文类型：学位论文

【摘要】：随着互联网信息技术和电子商务产业的快速发展,线上购物成为一种方便、快捷、有吸引力的购物方式,得到了数以十亿计的网络用户的关注。其中,服饰类商品在电商行业中具有十分重要的地位。通常,网络用户,尤其是女性消费者每天会花费几个小时来浏览、搜索和选择满足她们需求的服饰。因此,基于计算机视觉的服饰搜索服务具有很大的商业价值。然而,服饰类购物图像通常拍摄于自然户外场景,且由时尚模特穿着来进行展示,这些特性使针对服饰类商品的视觉搜索成为一个极具挑战性的课题。本文对服饰类商品图像进行研究,从中提取出服饰来增强视觉搜索的质量,主要内容如下：第一,提出了一种结合姿势检测和区域比较的服饰提取算法,其使用超像素分割算法将图像分成一系列区域,利用姿势检测定位服饰的大致区间,并将两者结合确定服饰的种子区域。然后,本文结合位置信息和HSV颜色特征来计算区域间的相似度,利用加权平方误差和来构造目标函数,将服饰分割问题转化为通过迭代计算和重分配区域类别来最小化目标函数。实验结果表明算法快速且具有鲁棒性,能够自动有效地进行服饰提取。第二,提出了基于衣物属性和人体结构的服饰提取优化。对于区域比较算法来说,前景可能包含多个不连通部分,本文利用区域的大小和位置对其进行重要性建模。考虑到服饰的完整性,被服饰区域包围的像素也应该为服装,因此我们将服饰内部像素分配为前景类别。本文还提出了一种基于最大后验概率的躯干模型,并利用躯干位置优化没有附着服饰的部位,如头部和下肢。最后,采用GrabCut算法对服饰提取结果进行像素级别优化。实验结果表明该算法能有效地提升服饰提取性能,适用于广泛的服饰图像。
[Abstract]:With the rapid development of Internet information technology and e-commerce industry, online shopping has become a convenient, fast and attractive way of shopping, which has attracted the attention of several 1 billion network users. Clothing products play a very important role in the e-commerce industry. In general, Internet users, especially female consumers, spend several hours a day browsing, searching and selecting clothing that meets their needs. Clothing search services based on computer vision have great commercial value. However, clothing shopping images are usually taken in natural outdoor scenes and displayed by fashion models. These characteristics make the visual search for clothing products become a very challenging subject. This paper studies the image of clothing commodities, and extracts clothing to enhance the quality of visual search. The main contents are as follows: first, A dress extraction algorithm combining posture detection and region comparison is proposed, in which the image is divided into a series of regions by using the hyperpixel segmentation algorithm, and the approximate range of clothing is located by posture detection. Then, combining the position information and HSV color features to calculate the similarity between the regions, the weighted square error sum is used to construct the objective function. The problem of clothing segmentation is transformed to minimize the objective function by iterative computation and redistribution of region categories. Experimental results show that the algorithm is fast and robust, and can automatically and effectively extract clothing. Second, The clothing extraction optimization based on clothing attributes and human structure is proposed. For the region comparison algorithm, the foreground may contain multiple disconnected parts. This paper uses the size and location of the region to model its importance. Considering the integrity of the dress, the pixels surrounded by the dress area should also be garments. In this paper, we propose a trunk model based on the maximum posteriori probability, and use the trunk position to optimize the non-attached parts, such as the head and lower limbs. The GrabCut algorithm is used to optimize the pixel level of the clothing extraction results. The experimental results show that the algorithm can effectively improve the performance of clothing extraction and is suitable for a wide range of clothing images.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】