结合标签优化和语义分割的服装图像检索

针对缺少民族服装语义标签、局部特征繁杂等因素导致少数民族服装图像检索准确率低的问题,提出一种结合标签优化和语义分割的服装图像检索方法.首先基于自定义的少数民族服装通用语义标签和民族服装语义标签,构建视觉风格分析概率模型进行标签优化;然后在全卷积网络结构基础上加入侧分支网络和全连接条件随机场,结合带有标注对和优化语义标签的训练图像对待检图像进行语义分割;最后采用多任务的深度监督哈希算法将语义分割结果哈希映射为二进制码,通过相似度计算对少数民族服装图像进行检索并输出结果.在构建的少数民族服装图像集上的实验结果表明,该方法能够有效地提高少数民族服装图像语义分割和检索的准确率.

Clothing Image Retrieval by Label Optimization and Semantic Segmentation

To address the problem of low accuracy of minority clothing image retrieval that lacks semantic labels and local feature complexity for minority clothing,a clothing image retrieval method using label optimization and semantic segmentation is proposed for various minority clothing.Firstly,a probabilistic model based on general semantic labels and minority semantic labels was constructed to analyze the visual style and optimize the defined semantic labels for minority clothing.Secondly,a side-branch network and a fully-connected CRF were added to the basis of the full convolutional network structure.Then,combining the training image with annotation pairs and optimized semantic labels,the query image was semantically segmented.Finally,a deep supervised hashing algorithm based on multi-task was used to encode the semantic segmentation result to a binary code,and the retrieval results of minority clothing image were output by similarity calculation.The proposed method is validated on the dataset of minority clothing.Experimental results show that the proposed method can effectively improve the accuracy of semantic segmentation and minority clothing image retrieval.