Low-resolution person recognition using image downsampling
Low-resolution person recognition using image downsampling
复制标题
DOI：
10.23919/mva.2017.7986904
复制
发表时间：
2017-05
期刊：
影响因子：
--
通讯作者：
K. Obara;Hiroki Yoshimura;Masashi Nishiyama;Y. Iwai
中科院分区：
文献类型：
--
作者：
K. Obara;Hiroki Yoshimura;Masashi Nishiyama;Y. Iwai
我们提出了一种通过使用分辨率推断对图像进行重新采样来识别个体的新方法，以提高人员识别的性能。图像的分辨率会动态改变面部或全身的外观，而低分辨率会降低人物识别的性能。为了克服分辨率问题，我们需要在识别之前使用重采样技术对图像的大小进行充分归一化。在我们的初步实验中，我们观察到对高分辨率图像进行下采样以适应低分辨率图像会提高识别性能，而对低分辨率图像进行上采样以适应高分辨率图像会降低性能。因此，所提出的方法通过比较查询图像和目标图像推断的分辨率，将下采样技术应用于更高分辨率的图像。我们证明，我们的方法大大提高了在人为降级为低分辨率的公开可用数据集 Multi-PIE 和 CUHK01 上的人物识别性能。
We propose a novel method for identifying individuals by resampling images using resolution inference to increase the performance of person recognition. The resolution of images dynamically changes the appearances of faces or whole bodies and low resolution decreases the performance of person recognition. To overcome the problem of resolution, we need to adequately normalize the sizes of images using resampling techniques before identification. In our preliminary experiment, we observe that downsampling high-resolution images to adjust to low-resolution images increases the identification performance while upsampling low-resolution images to adjust to high-resolution images decreases performance. The proposed method thus applies a downsampling technique to images of higher resolution by comparing the resolutions inferred for query and target images. We demonstrate that our method substantially improves the performance of person recognition on the publicly available datasets Multi-PIE and CUHK01 artificially degraded to low resolution.