2.3 Image extraction process
2.3.4 Texture and transform domain feature-based approaches
The palmprint image can be regarded as a textural image, comprising of the principal lines, wrinkles, ridges and other information. Palmprint recognition based on texture coding usually involves small feature size, fast matching speed and high accuracy in identification. The Fourier, wavelet, Gabor and Gabor wavelet transforms are all techniques that rely on the texture of the palm to extract features.
Li et al. [69] introduced a novel method to extract the features by transforming a palm image from a spatial domain to a frequency domain by employing the two-dimensional Fourier transform, following which the feature extraction. The researchers pointed out that the proposed approach shows high performance in terms of recognition rate and efficiency related to the palmprint database, as well as low time consumption.
Chen et al. [15, 16] proposed the contourlet transform, which is a novel two-dimensional extension of the wavelet transform, to be used to extract invariant palm features employing multiscale and directional filter data. The similarity between palmprint feature vectors can be measured using the AdaBoost classifier.
Figure 2.8: Example of the palmprint features extracted by 2D Gabor filtering [71]: (a) preprocessed palmprint image; (b) real part of texture palmprint image; (c) imaginary part of texture palmprint features.
Choge et al. used 2D-DCT to extract features from palmprints derived from the small blocks of the segmented part at the centre of the palm image [51]. The standard deviations of the 2D-DCT coefficients of each small patch are used as its features. This approach can minimize palmprint feature size in an effective way, but is suitable only for a small- scale database. Badrinath et al. [5] proposed another version with 1D-DCT coefficients of adjacent rectangular regions of variable size and orientation to represent palmprint images. The matching of the binary features of two palmprints is then conducted employing the Hamming distance, while the nearest neighbour approach is applied for classification.
Among the palm texture extraction methods, Gabor filter is more often employed to extract local direction features. Zhang et al. [122] suggested an online palmprint recognition system. In the pre-processing of a palmprint image, a low-pass filter and boundary tracking technique are adopted. The authors employed a 2D Gabor filter to convolve the palmprint image and the phase information of the filter responses is encoded as bitwise palm features. The normalized Hamming distance was employed as a similarity metric with regard to two 2048-dimensional texture feature vectors. Subsequently, this approach can represent the feature well, but as a result of using only a Gabor filter of 45◦ and disregarding other directions, it provides highly correlated features from different palmprint images.
Kong [55] proposed a new method for palmprint identification that employs low-resolution images and texture-based feature extraction. A palmprint is dealt with as a textural image and an adjusted Gabor filter is used to obtain information which is relevant to the texture. In addition, the Hamming distance is employed to assess the effectiveness of this particular process.
To improve performance, Kong et al. [52] proposed another approach termed the fu- sion code method, which convolves pre-processed palmprint images to encode the phase of the filter responses from a bank of Gabor filters in four orientations with the purpose of calculating four palm codes. It should be noted that, regarding the fusion rule, the four palm codes are combined to create the fusion code. Therefore, it is established that in this case the fusion code is independent of the brightness and contrast that is apparent in palm- print images. The similarity of two fusion codes is computed employing their normalized Hamming distance. Figure 2.8 represents an instance of the information extracted from a palmprint image using a 2D Gabor filter.
LBP is a widely used method in various application and a number of papers on LBP have been published. For example, Michael has proposed an innovative touchless palm- print identification system [75]. Hand tracking and palmprint ROI extraction technique are
employed to track and capture a user’s palmprint in a real-time video stream. Here, the LBP algorithm is employed for palmprint feature extraction. Classification is accomplished by applying a modified Probabilistic Neural Network (PNN). In [134], it is the first time that the LBP and LTP descriptors are applied to the energy or direction representations of palmprint extracted by the modified finite radon transformation (MFRAT), which to reduce the noise interference.
Recently a great progress has been made in LBP research. In [36] the hierarchical mul- tiscale LBP is a presented. It is an approach that improves the performance by extracting the information from the non-uniform bins. First, the LBP descriptor is extracted with a larg radius, and then, the counterpart LBP of those non-uniform patterns is extracted by a smaller radius. Compared with multi-scale LBP, their proposed technique enables more than 1% improvement. In another approach, Guo [32] proposed a novel collabora- tive representation model with a hierarchical multiscale local binary pattern (HM-LBP). The discriminative information from non-uniform palmprint patterned is extracted, and its dimension is reduced by PCA. This method has achieved excellent performance in both effective recognition accuracy and speed, which is able to fit for the real-time palmprint recognition system.
One of the successful extensions to the basic LBP operator is called Uniform LBP (ULBP) where the [98], ROI palmprint is decomposed into sub-bands with complex di- rectional wavelet coefficients up to 3-levels; each sub-band is divided into sub-blocks; and ULBP histogram of each sub-block is calculated. Then ,fisher linear discrimination is used reduce the size of the dimensionality.