A Study on Rotation and Scale Invariant Pattern Recognition Methods in Image Processing

Abstract

A major goal of the pattern recognition is to find methods that could provide some distortion-invariant recognition, i.e. the detection process has low sensitivity to the deformations of input objects. Typically, three deformations are considered: position, rotation, and scale changes of the input. In this thesis, the emphases of the research are two invariances: rotation invariant texture classification and scale invariant pattern recognition. (1) Rotation-invariant texture classification For rotation-invariant texture classification, we have proposed two methods. And a method for orientation estimation is introduced based on phase congruency. (1-1) Rotation-invariant texture classification Using Gabor wavelet The Gabor representation has been shown to be optimal in the sense of minimizing the joint two-dimensional uncertainty in space and frequency. And Gabor wavelet can be used to decompose an image into multiple scales and multiple orientations. Two group features i.e. the global feature vector and local feature matrix, can be constructed by the mean and variance of the Gabor filtered image. Global feature vector is rotation invariant, and local feature matrix can be adjusted by a circular shift operation to rotation invariant so that all images have the same dominant direction. By the two group features, a discriminant can be found to classify rotated images. In the primary experiments, comparatively high correct classification rates were obtained using a large sample sets with 1998 rotated images of 111 Brodazt texture classes. (1-2) Rotation-Invariant Texture Classification Using Circular Gabor Wavelets We present a new and effective method for rotation-invariant texture classification based on the circular Gabor wavelets. Two group features can be constructed by the mean and variance of the circular Gabor filtered images, and rotation invariant. By the two group features, a discriminant can be found to classify rotated images. The proposed method is evaluated on three public texture databases: Brodatz, CUReT and UIUCTex. The experimental results, based on different testing data sets, show that the proposed method has comparatively high correct classification rates not only on the rotated images, but also on the images under different illuminations and viewing directions. The proposed method is robust to additive white noise. (1-3) Orientation estimation from phase congruency The problem of orientation estimation is basic to many tasks in machine vision and image processing. A new approach for orientation estimation is proposed based on the phase congruency in radon domain. Here, the image principle direction is defined as the orientation of the image, which has the maximum of phase congruency of variance. The performance of this technique is determined by conducting simulation experiments on two sets of images, containing military targets and textures, respectively. (2) Scale Invariant Pattern Recognition For scale invariant pattern recognition, a novel scale and shift invariant pattern recognition method is proposed to improve the discrimination capability and noise robustness by combining the bidimensional empirical mode decomposition with the Mellin radial harmonic decomposition. The flatness of its peak intensity response versus scale change is improved. This property is important, since we can detect a large range of scaled patterns using a global threshold. Within this range, the correlation peak intensity (CPI) is relatively uniform with a variance below 20%. This proposed filter has been tested experimentally to confirm the result from numerical simulation for cases both with and without input white noise.