Analysis of adaptability of deep features for verifying blurred and cross-resolution images

Abstract

Employing Convolutional Neural Networks (CNN) to extract deep features from facial images for the task of recognition, identification and verification is well established. However, features extracted using CNNs have not been thoroughly studied for cross-resolution and blurred face verification. In this paper, we investigate the effectiveness of CNN features, that are primarily trained for matching high resolution images, to verify a pair of images constructed from a high and a low resolution face images. To perform this task, we degrade the image quality of the probe by artificially blurring and down sampling them, before it is passed to the CNN to be verified against high-resolution gallery image. After thorough experimental analysis, we present a pipeline which successfully improves upon the results obtained by raw CNN features, without any prior information of the quality of the degraded probe image. Using this pipeline, we show that the proposed system leads to improving verification accuracy in LFW and CMU-PIE datasets.