Feature extraction using discrete cosine transform and discrimination power analysis with a face recognition technology

Abstract

Discrete cosine transform (DCT) is a powerful transform to extract proper features for face recognition. After applying DCT to the entire face images, some of the coefficients are selected to construct feature vectors. Most of the conventional approaches select coefficients in a zigzag manner or by zonal masking. In some cases, the low-frequency coefficients are discarded in order to compensate illumination variations. Since the discrimination power of all the coefficients is not the same and some of them are discriminant than others, so we can achieve a higher true recognition rate by using discriminant coefficients (DCs) as feature vectors. Discrimination power analysis (DPA) is a statistical analysis based on the DCT coefficients properties and discrimination concept. It searches for the coefficients which have more power to discriminate different classes better than others. The proposed approach, against the conventional approaches, is data-dependent and is able to find DCs on each database. The simulations results of the various coefficient selection (CS) approaches on ORL and Yale databases confirm the success of the proposed approach. The DPA-based approaches achieve the performance of PCA/LDA or better with less complexity. The proposed method can be implemented for any feature selection problem as well as DCT coefficients. Also, a new modification of PCA and LDA is proposed namely, DPA–PCA and DPA–LDA. In these modifications DCs which are selected by DPA are used as the input of these transforms. Simulation results of DPA–PCA and DPA–LDA on the ORL and Yale database verify the improvement of the results by using these new modifications.

Introduction

Researches in the field of face recognition have become an attractive subject in the recent years. Activities in this field come from its applications in different areas such as security and surveillance, commercial and law enforcement. Ability for implementation in real time has intensified the attention to this field. Although research in the field of face recognition is active over 30 years and considerable successes in face recognition systems have been achieved, still there are some unsolved problems in it. Illumination variation, rotation and facial expression are the basic existing challenges in this area.

A wide variety of approaches have been developed by researchers for face recognition problem (see [1], [2] for a survey). From one point of view, these various approaches are categorized into two general groups, namely feature-based and holistic approach. In feature-based approaches, shapes and geometrical relationships of the individual facial features including eyes, mouth and nose are analyzed. But in holistic approaches, the face images are analyzed as two-dimensional holistic patterns. Feature-based approaches are more robust against rotation, scale and illumination variations, but their success depends on the facial feature detection. Due to difficulties in facial feature detection, holistic approaches are considered much more frequently than feature-based approaches [1].

Extracting proper features is crucial for satisfactory design of any pattern classifier system. In this way, two types of discrete transforms, statistical and deterministic, have been widely used for feature extraction and data redundancy reduction. The basis vectors of the statistical transforms depend on the statistical specification of the database and different basis vectors are possible for different databases. Deterministic transforms have invariant basis vectors which are independent of the database. Although statistical transforms have a great ability to remove correlation between data, they have high computational complexity. Also computation of the basis vectors for each given database is needed.

Among statistical approaches, principal component analysis (PCA) and linear discriminant analysis (LDA) are two powerful statistical tools for feature extraction and data representation. Kirby and Sirovich [3] were the first to employ Karhunen–Loeve transform (KLT) to represent facial images. Then Turk and Pentland [4]developed a PCA based approach namely “Eigenface”. PCA extracts features that are most efficient for representation, which may not be very useful for classification. Etemad and Chelappa [5], Belhumeur et al. [6] and Zhao et al. [7] then proposed the LDA “Fisherface” method to extract features that are most efficient for classification. Because of some limitations of the PCA and LDA, a variety of modifications have been proposed by authors. For more details see [8], [9], [10], [11], [12], [13], [14], [15], [16].

Proportional advantages of deterministic transforms make them an interesting class of feature extraction approaches. Discrete Fourier transform (DFT) [17], wavelet transform [18], [19], [20] and discrete cosine transform (DCT) [21], [22], [23], [24], [25], [26], [27], [28] are the important approaches of this class. Some special properties of DCT make it a powerful transform in image processing applications, including face recognition. DCT is very close to the KLT and has a strong ability for data decorrelation [26]. There are fast algorithms for DCT realization, which is not the case for KLT.

Combination of statistical and deterministic transforms constructs a third type of feature extraction approaches with both advantages. In this type, DCT reduces the dimension of data to avoid singularity and decreases the computational cost of PCA and LDA. Chen et al. [29] showed that using the PCA or the LDA in the DCT domain yields the same results as the one obtained from the spatial domain. Various combinations of the DCT, PCA and LDA have been surveyed by Pnevmatikakis and Polymenakos [30]. Ramasubramanian and Venkatesh [22] used a combination of the DCT, PCA and the characteristics of the human visual system for encoding and recognition of faces.

After applying the DCT to an image, some coefficients are selected and others are discarded in data dimension reduction. The selection of the DCT coefficients is an important part of the feature extraction process. In most of the approaches which utilize the DCT, not enough attention was given to coefficients selection (CS). The coefficients are usually selected with conventional methods such as zigzag or zonal masking. These conventional approaches are not necessarily efficient in all the applications and for all the databases. Discrimination power analysis (DPA) is a novel approach which selects features (DCT coefficients) with respect to their discrimination power. DPA utilizes statistical analysing of a database, associates each DCT coefficient discrimination to a number, and generates a CS mask. Our proposed feature extraction approach is database dependent and is able to find the best discriminant coefficients for each database.

The rest of the paper is organized as follows: Section 2 is devoted to the feature extraction approach. The whole procedure of our approach is depicted in Section 3. Experimental results and discussion are presented in Section 4, and finally Section 5 includes conclusion.