부산대학교 도서관

Multi-modal images are commonly used in the field of medicine for anomaly detection, for example CT/MRI images for tumour detection. Recently, thermal imaging has demonstrated its potential for detection of anomalies (e.g., water stress, disease) in plants. In biology, multi channel imaging systems are now becoming available which combine information about the level of expression of various molecules of interest (e.g., proteins) which can be employed to investigate molecular signatures of diseases such as cancer or their subtypes. Before combining information from multiple modalities/channels, however, we need to align (register) the images together in a way that the same point in the multiple images obtained from different sources/channels corresponds to the same point on the object (e.g., a particular point on a leaf in a plant or a particular cell in a tissue) under observation. In this thesis, we propose registration methods to align multi-modal/channel images of plants and human tissues. For registration of thermal and visible light images of plants we propose a registration method using silhouette extraction. For silhouette extraction, we propose a novel multi-scale method which can be used to extract highly accurate silhouettes of diseased plants in thermal and visible light images. The extracted silhouettes can be used to register plant regions in thermal and visible light images. After alignment of multi-modal images, we combine thermal and visible light information for classification of water deficient regions of spinach canopies. We add depth information as another dimension to our set of features for detection of diseased plants. For depth estimation, we use disparity between stereo image pair. We then compare different disparity estimation algorithms and propose a method which can be used to obtain not only accurate and smooth disparity maps but also less sensitive to the acquisition noise. Our results show that by combining information from multiple modalities, classification accuracy of different classifiers can be increased. In the second part of this thesis, we propose a block-based registration method using mutual information as a similarity measure for registration of multi-channel fluorescence microscopy images. The proposed block-based approach is fast, accurate and robust to local variations in the images. In addition, we propose a method for selection of a reference image with maximal overlap i.e., a method to choose a reference image, from a stack of dozens of multi-channel images, which when used as reference image causes minimum amount of information loss during the registration process. Images registered using this method have been used in other studies to investigate techniques for mining molecular patterns of cancer. Both the registration algorithms proposed in this thesis produce highly accurate results where the block-based registration algorithm is shown to be capable of registering the images up to sub-pixel accuracy. The disparity estimation algorithm produces smooth and accurate disparity maps in the presence of noise where commonly used disparity estimation algorithms fail to perform. Our results show that by combining multi-modal image data, one can easily increase the accuracy of classifiers to detect anomalies in plants, which helps to avoid huge losses due to disease or lack of water at commercial level.