Interpolation of low resolution images for improved accuracy in human face recognition

Abstract

This paper presents the effect of interpolation schemes, namely, nearest-neighbor, bilinear, and bicubic, when applied to low-resolution images as a preprocessing step for improving the recognition rate in human face recognition system. Three feature extraction methods are used, namely, Local Binary Pattern, Discrete Wavelet Transform, and Block-Based Discrete Cosine Transform, with and without interpolation for comparison purpose. The experiments are conducted on the ORL database. Bicubic and bilinear improve the recognition rate of low resolution images considerably. NTRODUCTION Human face recognition (FR) constitutes a very active area of research for the last two decades and many challenges have been addressed, such as, illumination variation, pose variation, local shadow etc. however, in real-world applications, such as, video surveillance cameras, where it is often difficult to obtain good quality recordings of observed human faces, face recognition is still challenging task. One important factor evaluating the quality of the recordings is the resolution of the images. Human face recognition from low resolution (LR) images has been tackled by many researchers [1]- [4]. In this paper we have utilized three different interpolation techniques as a preprocessing step for improving the quality of lowresolution images before feature extraction as will be discussed in section 3. It is pointed out that image-based (holistic) approach is superior for LR face recognition [1]. Thus, we have chosen three effective feature extractors, namely, Local Binary Pattern (LBP) [5], Discrete Wavelet Transform (DWT) based PCA [6-8], and Block-Based Discrete Cosine Transform (BBDCT) [9] for our experiments and the classifier used is Euclidean classifier.This paper is organized as follows. Section 2 is an overview of the face descriptors that will be used. Interpolation techniques illustrated In Section 3, and experiments and results are discussed in section 4 and concluding remarks are given in section 5. FEATURE EXTRACTIONThis section outlines the feature extraction schemes that are used in this paper: LBPThis operator labels the pixels of an image by thresholding the 3×3 neighbourhood of each pixel with the center value and considering the result as a binary number (see Figure 1). Then the histogram of the labels can be used as a texture descriptor. Using circular neighbourhoods and bilinearly interpolating the pixel values allow any radius and number of pixels in the neighbourhood. Further extension of LBP is to use uniform patterns. A uniform pattern has at most two bitwise transitions from 0 to 1 or vice versa when the binary string is considered circular. In this work, we adopt a LBP operator , , that selects uniform patterns in (8,1) neighbourhood; where the radius is one with 8 neighbors. It is observed that uniform patterns account for nearly 90% of all patterns in the (8, 1) [5]. We select Chi square statistic ( ) as the dissimilarity measure for histogramWhere S and M are two LBP histograms; , is the histogram of reg...