New Texture Descriptors

Inspired by the huge success of the Local Binary Pattern (LBP) method, we have also proposed new local texture descriptors, including the blur-insensitive Local Phase Quantization (LPQ) method and the descriptor based on Weber's law (WLD) that have provided state-of-the-art performance e.g. in face recognition and texture classification problems. Both LPQ and WLD are related but complementary to the LBP method.

lpq1.jpg A simplified procedure for computing LPQ.

Local Phase Quantization (LPQ)

Fourier phase spectrum can be shown to be invariant to image blurring with centrally symmetric point spread functions (PSF) at such frequencies where the Fourier transform of the PSF is positive. In many cases real image blur can be approximated by Gaussian blur, linear motion blur, or disk-shaped blur that all have centrally symmetric PSFs. Local Phase Quantization (LPQ) is a novel texture descriptor that utilizes the blur invariance property of the phase spectrum. It is based on binary coding of the quantized Fourier phase computed locally around each pixel. Because of the limited resolution of the local phase information, the descriptor obtained is not completely invariant to the blur, but it has been experimentally verified that it is still highly insensitive to moderate blurring. Also, for sharp images LPQ has proved to be an extremely powerful descriptor. In comparative studies it has outperformed other texture descriptors such as LBP. Applications of LPQ include face recognition and medical image analysis.

Blurred face images.

Matlab implementation of LPQ is available here.

More information about LPQ can be found from the following publications:

For more publications, see LBP Bibliography.

Weber's Law Descriptor

In collaboration with the Institute of Computing Technology of the Chinese Academy of Sciences, we proposed a simple, yet very powerful and robust local descriptor, called the Weber Local Descriptor (WLD). It is based on the fact that human perception of a pattern depends not only on the change of a stimulus (such as sound, lighting) but also on the original intensity of the stimulus. Specifically, WLD consists of two components: differential excitation and orientation. The differential excitation component is a function of the ratio between two terms: one is the relative intensity differences of a current pixel against its neighbors; the other is the intensity of the current pixel. The orientation component is the gradient orientation of the current pixel. For a given image, we use the two components to construct a concatenated WLD histogram. Experimental results on texture analysis and face detection problems provided excellent performance. Recently, we combined LBP and WLD for the segmentation of dynamic textures and provided very good segmentation results compared to the state-of-the-art.

Illustration of the computation of the Weber Law Descriptor.

Matlab implementation of WLD is available here.

More information about WLD can be found from the following publications:

CMV/Research/NewTextureDescriptors (last edited 2011-11-22 12:09:35 by WebMaster)