Research On Scaling And Rotation Invariant Analysis Methods For Images Recognition

Geometric transformation invariant image recognition is a very active research area in image processing and pattern recognition with many applications such as machine vision, remote sensing, medical imaging etc. There are four stages to image recognition including image input, image preprocessing, extracting features from image and classification of image. The removal of noise is an important and traditional problem of image preprocessing. There are many works on the restoration of images corrupted by noise. Texture analysis plays an important role in machine vision and image processing; rotation invariant texture classification and texture orientation estimation have been of particular interest. Manifold learning is a popular way to classify images under unconstrained condition; it is also an important issue in data mining. Recently, manifold learning has been the focus of geometric transformation invariant image recognition due to the fact that geometric transforms are important factors of constrained condition.In this dissertation, the restoration of images corrupted by noise is concerned, a new image filtering algorithm for removing salt-and-pepper impulse noise and a nonlinear filtering algorithm using probability statistic and local main texture direction analysis are proposed. We also focus on the issue of scaling and rotation invariant image recognition and present two approaches, and then rotation invariant analysis and orientation estimation methods for texture classification are considered, two rotation invariant methods for texture classification and texture orientation estimation are introduced. We finally discuss manifold learning and propose an improved isometric map (ISOMAP) method.The main contributions of this dissertation are summarized as follows:(1) A nonlinear filtering algorithm using probability statistic and main texture direction analysis is proposed. In this algorithm Radon transform is utilized to determine texture direction probability density distributions of local areas of images and then a probability statistic model is applied to estimate the middle pixel's gray value according to its neighbour pixels. The superiority of this method is the high ability of denoising and preserving edges and details of images corrupted by pulse noise and Gaussian noise. Unlike some recent algorithms only for removing salt-and pepper impulse or Gaussian noise, this method is applicable to images contaminated by the mixture of pulse noise and Gaussian noise. We also present a new image filtering algorithm using a double noise detector and edge-preserving regularization function. The proposed filter has a two-stage scheme: detecting noise and removing noise. In order to improve accurate rate of noise detection, noise candidates identified with the noise detection algorithm of the adaptive median filter are judged again by local fuzzy membership function, and a convex objective function composed of data-fidelity term and edge-preserving regularization function is employed to deal with noise candidates. The input of edge-preserving regularization function is adaptively selected to take full advantage of local features of images. The image corrupted by noise is restored successfully as the convex objective function gets its minimum. This method yields better performance in removing noise and preserving the details and edges of the image in comparison with some recent methods, even at a very high noise level (>70%) details and edges of the original image are preserved very well with our method.(2) A new scaling and rotation invariant analysis method for object recognition is proposed, the Radon transform is utilized to project the image onto projection space to convert the rotation of the original image to a translation of the projection in the angle variable and the scaling of the original image to a scaling of the projection in the spatial variable together with an amplitude scaling of the projection, and then the Fourier–Mellin transform is applied to the result to convert the translation in the angle variable and the scaling in the spatial variable as well as the amplitude scaling of the projection to a phase shift and an amplitude scaling, respectively. In order to achieve a set of completely invariant descriptors, a rotation and scaling invariant function is constructed. A k-nearest neighbors'classifier is employed to implement classification. Theoretical and experimental results show the high classification accuracy of this approach as a result of using the rotation and scaling invariant function instead of image binarization and normalization, it is also shown that this method is relatively robust in the presence of white noise(3) A new rotation invariant analysis approach to texture image classification is described. In the proposed approach, the Radon transform is employed to project the texture image to be analyzed onto projection space, and then bispectrum analysis approach is applied to implement texture rotation invariant classification and calculation of the angle which the texture image was rotated by. The texture patterns extracted from the Radon projection space of the texture image are global and less sensitive to white noise. Since the bispectrum results of Gaussian noise with zero mean are all zero, the proposed approach is less sensitive to Gaussian noise. Experimental results show the high classification accuracy of this method. It is also shown that this proposed approach is relatively robust in the presence of noise. We also present a novel rotation invariant analysis and orientation estimation method for texture classification. In the proposed method, the Radon transform is utilized to project the texture image onto projection space to convert the rotation of the original texture image to a translation of the projection in the angle variable, and then correlation analysis technique is employed to the results to classify the texture image and estimate the orientation of the texture image. This method is robust in the presence of white noise and yields better performance in orientation estimation and texture classification in terms of correct classification percentages compared with recent methods.(4) The success of ISOMAP depends on being able to choose a neighbourhood size. An improved ISOMAP algorithm is proposed. In the improved Isometric mapping algorithm, the average shortest distance of the data set is utilized to determine a range of neighbourhood sizes and then an approach to eliminating short-circuit edges is presented to improve the degree of noise tolerance of the ISOMAP algorithm. This method is relatively robust in the presence of high noise levels and the computational cost is lower than some recent methods.