Current Issue Cover


摘 要
目的 在传统的词袋模型图像搜索问题中,许多工作致力于提高局部特征的辨识能力。图像搜索得到的图像在细节部分和查询图像很像,然而有时候这些图像在语义层面很却差别很大。而基于全局特征的图像搜索在细节部分丢失了很多信息,致使布局相似实则不相关的图像被认为是相关图像。为了解决这个问题,本文利用深度卷积特征来构建一个动态匹配核函数。 方法 利用这个动态匹配核函数,在鼓励相关图像之间产生匹配对的同时,抑制不相关图像之间匹配对的个数。该匹配核函数将图像在深度卷积神经网络全连接层最后一层特征作为输入,构建一个动态匹配核函数。对于相关图像,图像之间的局部特征匹配数量和质量都会相对增强。反之对于不相关的图像,这个动态匹配核函数会在减少局部特征匹配的同时,降低其匹配得分。结果 本文从数量和质量上评估了提出的动态匹配核函数,提出了两个指标来量化匹配核函数的表现。基于这两个指标,本文对中间结果进行了分析,证实了动态匹配核函数相比于静态匹配核函数的优越性。最后,本文在5个公共数据集进行了大量的实验,在对各个数据集的检索工作中,得到的平均准确率从85.11%到98.08%均高于此领域的同类工作。实验表明了本文提出的方法是有效的,并且其表现优于当前这一领域的同类工作。
Dynamic Match Kernel for Image Retrieval

Hong Rui()

Objective In conventional bag-of-visual-words (BoW) based image retrieval, lots of attentions are paid to enhancing discrimination of local features. The images returned are similar in details with query image, while it is dramatically different from a semantic perspective. However, image retrieval based on global descriptors will loss many details between images, which will result in that the images similar in layout while different in semantic are considered as relevant pairs. To address the problem, we propose a dynamic match kernel which take advantage of convolutional neural network. Method The proposed match kernel will stimulate the feature matches between near-duplicate images and filters the matches between irrelevant images. We extract features from the last fully connected layer in convolutional neural network as the input for dynamic match kernel, then a adaptive threshold are constructed for matching the local features. For relevant images, the threshold should large as much as it possible, so that the positive matches will be included more, and the vice versa. Result Additionally, we proposed two criteria to evaluate the effect of dynamic match kernel, and show its superiority over static match kernel in both quantity and quality of positive matches. Finally,extesive experiments have been done in 5 public datasets. The mAP obtained from searching data sets is from 85.11% to 98.08%, which is higher than other method in this field. The experiment demonstrate that the proposed method is effective and outperforms state of the art.