基于哈希的细粒度图像检索方法研究

The task of fine-grained image retrieval is to search images with multiple levels semantic relationship. Compared with category-level image, fine-grained image retrieval meet users’ requirement better. It is the new problem and trend in the field of content-based image retrieval. For large-scale image dataset, the project intends to study the fine-grained image retrieval method based on hashing. Firstly, feature extraction method based on parameter compression in deep convolutional neural network is studied, which can obtain high dimensional image feature with strong recognition ability. Secondly, topology preserve image hashing method based on bilinear projection is proposed, and the image hashing model based on aggregation and product quantization is constructed for local features. Combining global and local binary codes, a fine-grained image retrieval model based on hashing is proposed to preserve the global similarity and highlight the local difference. As a result, it can improve the precision and recall of images with the same semantic and similar semantic. Finally, binary codes weights computation method based on sequence pairwise is proposed. Even more, the weighted Hamming distance is computed to rerank the returned images with multiple levels semantic relationship. As a result, the large-scale fine-grained image retrieval is achieved with high precision and rapid speed.

细粒度图像检索的主要任务是搜索包含多重语义关系的图像。相对于同类图像检索，细粒度图像检索更符合用户需求，是基于内容图像检索研究的新问题和发展趋势。面向大规模图像库，本项目拟研究基于哈希的细粒度图像检索方法。首先，研究基于深度卷积网络参数压缩的特征提取方法，以获得辨识能力强的高维图像特征；其次，提出基于双线性映射的拓扑保持图像哈希方法，建立基于聚合和乘积量化的局部特征哈希模型，采用多特征倒排表结构联合全局和局部二值特征，研究全局相似局部差异的细粒度图像检索方法，提高语义相同和语义相似图像的检索准确率和召回率；最后，提出基于排序对的二值哈希码权重计算方法，建立基于加权汉明距离的图像重排模型，优化具有多重语义关系的图像排序结果，实现对大规模图像库快速准确得细粒度图像检索。

细粒度图像检索的主要任务是搜索包含多重语义关系的图像。相对于同类图像检索，细粒度图像检索更符合用户需求，是基于内容图像检索研究的新问题和发展趋势。面向大规模图像库，本项目研究了基于哈希的细粒度图像检索方法。首先，研究了基于深度卷积网络的特征提取方法，以获得辨识能力强的高维图像特征；其次，建立了基于顶端相似性保留的深度哈希模型，提出了局部相关的多特征关系图融合方法，联合全局和局部二值特征，提高语义相同和语义相似图像的检索准确率和召回率；最后，提出辅助点上下文建模的图像重排序方法，优化具有多重语义关系的图像排序结果，实现对大规模图像库快速准确得细粒度图像检索。