结合显著区域检测和手绘草图的服装图像检索

针对服装图像检索准确率和效率较低的问题,提出一种服装显著区域检测和手绘草图的服装图像检索方法。首先采用正则化随机漫步算法对输入的服装图像库进行视觉显著区域检测,并结合其边缘轮廓信息,得到服装显著边缘图像;其次,对输入的服装草图和服装边缘图像进行特征提取,得到服装草图和服装边缘图像各自的方向梯度直方图(HOG)特征;然后,通过计算服装草图特征和服装边缘特征的相似度,实现特征匹配;最后,按照特征匹配结果在服装图像库中检索与服装草图相似的服装图像,采用基于距离相关系数的重排序算法对其相似度进行排序并输出检索结果。结果表明,该方法提高了服装检索的准确率,具有较好的鲁棒性,检索准确率可达78.5%。

Clothing image retrieval by salient region detection and sketches

In order to solve the problems of unsatisfactory accuracy and low efficiency in the clothing image retrieval, a sketch based clothing image retrieval method by visual salient regions and re-ranking was proposed. Firstly, clothing salient edge map was obtained by saliency detection method with regularized random walks walking and the edge map. Then, histogram of oriented gradeient features of user sketches and the salient edge in clothing images were extracted, respectively, and the feature matching was achieved by similarity calculation between the input sketches and clothing images. Finally, the retrieval results were output in descending order according to the similarity. Using the re-ranking optimization based on distance correlation coefficients, final results were obtained. Experimental results show that the method can effectively provide clothing retrieval results and significantly improve accuracy and robustness comparison with other approaches. The accuracy ratio of the algorithm is higher than 78.5%.