Face recognition using part-based dense sampling local features

Abstract

For years, researchers have made great efforts to find an appropriate face representation for face recognition. A fusion strategy of Local Binary Pattern (LBP) and Gabor filters yields great achievements. LBP is good at coding fine details of facial appearance and texture, whereas Gabor features can encode facial shape and appearance over a range of coarser scales. Despite the great performance, this fusion representation suffers from low effectiveness and resolution variance. In this paper, we propose a novel representation strategy of face images which is fast and robust to resolution variance. We apply dense sampling around each detected feature point, extract Local Difference Feature (LDF) for face representation, then utilize Principal Component Analysis (PCA)+Linear Discriminant Analysis (LDA) to reduce feature dimension and finally use cosine similarity evaluation for recognition. We have utilized our proposed face representation strategy on two databases, namely self-collected Second Generation ID Card of China and Driver׳s License (SGIDCDL) database and public Facial Recognition Technology (FERET) database. Our experimental results show that the proposed strategy has good performance on face recognition with fast speed.

Introduction

Over the last few decades, face recognition has been an active research topic in computer vision and pattern recognition, and a lot of such systems emerged across various applications, such as video surveillance, access control, image retrieval and automatic log-on for personal computers or mobile devices. Face recognition is an easy task for human visual system, but it is quite challenging for automatic face recognition system due to the dramatic variations among the appearance of the same subject, which is caused by a lot of factors such as image resolution, illumination, expression, pose, occlusion, etc. These various visual complications deteriorate the performance of a face recognition system dramatically.

Automatic face recognition mainly involves three stages: detection, representation and classification. In the detection stage, the face is localized in an image. The representation stage refers to extracting specific features from a given face image. The classification stage deals with the final decisive process whether an unknown face image can be categorized into a target subject. Among the three main parts, face representation occupies a crucial position of a face recognition system. A critical point to the performance of a face recognition system is the face representation׳s discriminative power for various individuals and invariance to external environment changes, such as resolution and poses variation. A robust and effective face representation is indispensable to performance improvement.

Traditional face representations such as Local Binary Pattern (LBP) [1,2] and Gabor [3], [4] achieved good performance on face images with good quality. LBP is good at coding fine details of facial appearance and texture, and Gabor features encode facial shape and appearance over a range of coarser scales. A fusion strategy of LBP and Gabor can obtain promising performance for face recognition [20], [22], [23], [24]. But this strategy has the problem of complex computations. Meanwhile, the fusion of LBP and Gabor is sensitive to resolution variation.

In addition to efficient face representation, the performance of face recognition system relies on the metric techniques used to measure the similarity of two face representations. The recent frameworks mostly learn a dataset specific metric during the training stage, which can result in the improvement of the recognition performance. It is mainly because it learns the bias of the dataset and thus adapts to the specific distribution. In this paper, we show that our face representation can be used with a very simple face matching method that does not require complex metric learning algorithm. Once features are extracted from face images, we first use whitened Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) to project them into a low dimensional space, and then apply cosine similarity on the more compact features to measure the similarity between image pairs.

We experimentally evaluated our proposed face representation on two databases. The first one, which is composed of a subject’s high resolution identification (ID) photo and relative low resolution driver׳s license photo, is collected by us. The second one is the public FERET database [5]. Experimental results show that our method has very good generalization capability and outperforms many traditional methods both in performance and efficiency.

The main contributions of this paper are summarized as follows:

• We propose a novel face representation strategy for face recognition, which is effective and robust to resolution variation.

• Experimental results show that the proposed face representation can significantly improve face recognition performance without time consuming Gabor convolution.

The rest of the paper is organized as follows: Section 2 reviews the related work of face recognition. Section 3 describes details of the whole framework and focuses on the proposed new face representation strategy. Section 4 shows experimental results and Section 5 gives conclusion of this paper.