Abstract/Summary

The ability to automatically recognise, classify, and cluster images is important in many applied fields, including medical healthcare, astronomy, entertainment, sports, defense, etc. Thus, it is important to develop efficient algorithms that can be used to recognise images in different scenarios. In this thesis, three novel image-processing techniques based on content features are proposed. First, modified probabilistic neural networks (PNNs) are introduced for image classification based on various distance measures in probability space, in which the input to the model is the local binary pattern (LBP) histogram of images. Conventional PNNs have an input layer that computes the Euclidean distance of pairwise input features. The proposed modified PNN considers various probability distance measures for computing the distances of LBP histograms between images. It is shown that a PNN based on the Bhattacharyya distance measure is superior to other studied measures. Moreover, using the subset of uniform LBP features is generally better than using full LBP features. Secondly, a novel self-tuning spectral clustering technique was proposed for image classification application. Uniform and full-histogram LBP were used as features, while modified affinity matrices based on four distance measures were tested. The experimental results indicate that uniform LBP features generally achieved higher accuracy when compared to full-histogram LBP features. On the distance measure side, Bhattacharyya distance-based affinity matrices achieved higher accuracy than other distance measures, especially in terms of large image data sets. The disadvantage of supervised learning techniques is that they are limited to learning from labelled data sets, which are often expensive to obtain. A novel de- cision fusion framework is proposed by combining semi-supervised clustering with well-known image feature analysis methods in computer vision. Initially, image features are generated by applying the Gray level co-occurrence matrix analysis to the processed data sets transformed by Gabor, Laplacian, or Gaussian filters. Then, an on-line spherical k-means clustering technique guided by a minimum number of labelled data sets is used to train the base classifiers. The final decision of classification is produced by selecting the classifier that produces the max-cosine value among the baseline classifiers. Comparative experiments have been conducted to demonstrate that the proposed approaches are suitable for automatic recognition.