Article Preview
Top1. Introduction
Face recognition represents a longstanding field of research in pattern recognition. Its popularity is justified by its usefulness in terms of multiple applications areas ranging from video surveillance and security to human computer interaction. Indeed, the face recognition task is still an unresolved problem that hasn't been fully deciphered because faces are captured in wild from far away cameras under challenging illumination and acquisition conditions. This problem is defined as the Low Resolution Face Recognition (LR FR) where gallery face images are in High-Resolution (HR) while probe face images are in Low-Resolution (LR). From this perspective, most of the existing works focus on the mismatch of HR-LR resolution. To resolve such mismatch three standard categories are reported: (A) Super-Resolution (SR) where the probe image is up-sampled to the resolution of the gallery images (Wang & Tang, 2005); (B) Down-sampling which reduces the gallery image to the probe image and then matching step is applied images (Wang & Tang, 2005); (C) Unified features space is used to project high-resolution gallery images and low-resolution probe images in a common space (Li, Chang, Shan & Chen, 2009).
In this context, we introduce a new face recognition method composed of two phases: an off-line phase and an inference phase. In the off-line phase, we built the Kernel Partial Least Squares (KPLS) regression model to map the LR facial features to HR version. Next, we used the KPLS predictor in the inference phase to map HR features from LR features. We applied in both phases the Block-Based Discrete Cosine Transform (BBDCT) vector to enhance the description of facial features. Finally, we carried out the identity matching using the K-Nearest Neighbor (KNN) classifier.
The main contributions of this work can be sum up as follows:
- •
We adapted a KPLS regression algorithm to map HR features from LR features to handle both of Low as well as Very Low resolution face recognition problems.
- •
We extracted a representative feature vector using the BBDCT descriptor that combine the global and the local features to further encode the facial feature in the LR and VLR images.
The rest of the paper is organized as follows. Section 2 is dedicated to a brief overview of the related works. The proposed method is identified and detailed in section 3. The experimental results are reported in section 4. Section 5 presents the conclusion and provides new perspectives for future works.
TopIn facial recognition applications, all images in the gallery are usually in high-resolution. The degradation of the query image resolution leads to the problem of dimensional incompatibility between the gallery and the probe face images in the matching step. This problem is known as the low resolution face recognition (LR FR).
In literature, the LR FR methods can generally be divided into two categories. The first category includes direct methods that extract facial features directly from the LR images. The second category includes indirect methods which first improve the resolution of LR images then extract their features.