Pattern recognition techniques for texture retrieval and gene expression data analysis

Abstract

Pattern recognition techniques have been applied on various applications such as speech recognition, computer vision and bioinformatics. In this work, these techniques are applied to texture image retrieval and gene expression data analysis. For texture image retrieval, a multiscale directional filter bank (MDFB) is studied. Our results show that the filters for scale decomposition should be designed to avoid aliasing so as to prevent corruption in directional decomposition. On the other hand, we propose to pre-classify structured textures and random textures in order to optimize the overall retrieval accuracy. In addition, feature reduction and rotation-invariance issues are addressed. The MDFB extends pyramidal DFB (PDFB) by introducing additional scale decomposition on the finest scale. Although the modification can improve the radial frequency resolution for texture characterization, it increases the computational complexity substantially. To minimize the complexity, we propose a new implementation in which the DFB is applied prior to the additional scale splitting. With decimation, the redundancy of the MDFB can be reduced to that of the PDFB. Our experimental results find that there is a speed up of more than 30% without performance deterioration in texture retrieval. One of the major tasks in gene expression data analysis is to identify genes with similar functions. Usually, these genes co-express under a subset of conditions only. Therefore, we investigate biclustering techniques which can cluster rows and columns simultaneously. In this thesis, a new biclustering algorithm based on a multidimensional visualization technique known as parallel coordinate (PC) plots is proposed. This algorithm uses a 'merge-and-split' strategy to grow biclusters from the clusters discovered in the PC plots. The advantages of the proposed algorithm include low computational complexity (polynomial time), high noise robustness and high effectiveness in detecting biologically-relevant biclusters. The new biclustering algorithm is included in our software tool named as BiVisu together with the PC plots for interactive analysis. The PC plots allow human to interpret the bicluster data in an intuitive way. In particular, an exploratory approach is designed to adjust the parameters of the biclustering algorithm with the use of PC plots.