目标提取与哈希机制的多标签图像检索

发布时间：2019-03-15 15:34

【摘要】：目的哈希是大规模图像检索的有效方法。为提高检索精度,哈希码应保留语义信息。图像之间越相似,其哈希码也应越接近。现有方法首先提取描述图像整体的特征,然后生成哈希码。这种方法不能精确地描述图像包含的多个目标,限制了多标签图像检索的精度。为此提出一种基于卷积神经网络和目标提取的哈希生成方法。方法首先提取图像中可能包含目标的一系列区域,然后用深度卷积神经网络提取每个区域的特征并进行融合,通过生成一组特征来刻画图像中的每个目标,最后再产生整幅图像的哈希码。采用Triplet Loss的训练方法,使得哈希码尽可能保留语义信息。结果在VOC2012、Flickr25K和NUSWIDE数据集上进行多标签图像检索。在NDCG(normalized discounted cumulative gain)性能指标上,当返回图像数量为1 000时,对于VOC2012,本文方法相对于DSRH(deep semantic ranking hashing)方法提高2 4个百分点,相对于ITQ-CCA(iterative quantization-canonical correlation analysis)方法能提高3 6个百分点;对于Flickr25,本文方法比DSRH方法能提高2个左右的百分点;对于NUSWIDE,本文方法相对于DSRH方法能提高4个左右的百分点。对于平均检索准确度,本文方法在NUSWIDE和Flickr25上能提高2 5个百分点。根据多项评价指标可以看出,本文方法能以更细粒度来精确地描述图像,显著提高了多标签图像检索的性能。结论本文新的特征学习模型,对图像进行细粒度特征编码是一种可行的方法,能够有效提高数据集的检索性能。
[Abstract]:Objective Hash is an effective method for large-scale image retrieval. In order to improve the retrieval accuracy, hash codes should retain semantic information. The more similar the images, the closer the hash code should be. The existing methods first extract the features describing the whole image, and then generate hash codes. This method can not accurately describe the multiple objects contained in the image, which limits the accuracy of multi-label image retrieval. A hash generation method based on convolution neural network and target extraction is proposed. Methods A series of regions that may contain objects in the image are extracted firstly, then the features of each region are extracted and fused by using the depth convolution neural network, and a set of features is generated to depict each target in the image. Finally, the hash code of the whole image is generated. The training method of Triplet Loss is used to make the hash code retain semantic information as much as possible. Results Multi-label image retrieval was performed on VOC2012,Flickr25K and NUSWIDE datasets. In terms of the NDCG (normalized discounted cumulative gain) performance index, when the number of returned images is 1,000, the proposed method for VOC2012, is 24 percentage points higher than the DSRH (deep semantic ranking hashing) method. Compared with ITQ-CCA (iterative quantization-canonical correlation analysis) method), it could be increased by 36 percentage points. For Flickr25, this method can increase about two percentage points compared with DSRH method, and for NUSWIDE, this method can increase about 4 percentage points compared with DSRH method. For the average retrieval accuracy, this method can improve 25 percentage points on NUSWIDE and Flickr25. According to several evaluation indexes, it can be seen that the proposed method can accurately describe the image with finer granularity, and significantly improve the performance of multi-label image retrieval. Conclusion the new feature learning model in this paper shows that fine-grained feature coding is a feasible method to improve the retrieval performance of data sets.
【作者单位】：国防科学技术大学并行与分布处理重点实验室;
【基金】：国家自然科学基金项目(U1435219)~~
【分类号】：TP391.41

【相似文献】