本文针对现有服装图像特征描述方法不能较好的概括服装特征,致使服装图像检索方法检索性能较低,或依赖文本描述的问题,提出一种融合高层次全局服装特征和中层服装区块特征的特征描述方法。首先,将图像分割成含有相对独立语义信息的片段;再根据片段的纹理、几何和色彩特征将其聚类为具有强语义信息的服装区块;之后提取服装区块的几何分布和色彩特征,与服装图像的全局特征进行融合,构成多层次的服装特征描述向量。采用上述特征对服装进行描述,并以文本描述为输入进行检索实验,结果表明该方法能够有效提高实现对服装类别、穿着场合等信息进行检索。

Clothes retrieval methods based on text description(tags)are not satisfying in effectivenessandaccuracymainly because the tags are derived from subjective human description and the cognitive differences are unavoidable.Thus vision features based descriptions are introduced for better retrieval results.Current description methods,mostly using clothes images with single-layer(high or low)features,either fail to describe clothes effectively in retrieval application,or require text tags to narrow down the retrieval range.For the latter situation,tags still bring in the inaccuracy caused by text description.To omit the affect by text and improve the retrieval,a novel method combining high-layer global features and mid-layer blocks' features is promoted to realize retrieval only by images.The method is based on the global-to-local process human cognition.To obtain the global description of the clothes image in high-layer,the improved histograms of primary color and primary oriented gradients are used to describe the color and geometry of the clothes.To obtain the mid-layer semantic description,local features in low-level are abstracted and combined.Firstly,a clothing image is segmented into visually distinguished pieces with graph-based segmentation,hence each piece holding simplex semantic information different from its background.To describe the piece semantically,improved methods are used to generate the feature vector from the texture,geometry and color features.Secondly,a cluster method is adopted to combine the semantic pieces into blocks based on their visual characteristics.As the converging of the homogeneous semantic pieces,the combined blocks hold enriched semantic information of part of the clothes,containing shape,style,material and so on.The geometric distribution and color features of the blocks are abstracted to describe the block and these features are finally combined with the above-mentioned global features into the feature vector of the image,which is introduced into the retrieval for clothes.In experiment,text descriptions are used as input for the retrieval process,and the results show efficiency in retrieval of three different aspects,and especially high accuracy on search with classification and occasion,which prove the effectiveness and universality of our method.