Image Difference Visualization Based on Mutual Information

We propose algorithm for local image difference measurement for multimodal image data based on value of mutual information of both images. Algorithm works with registered gray-level images. Similarity measure takes into account entropy of compared images. Results of proposed algorithm can be used for image comparison and better difference localization. Introduction Multimodal data sets are nowadays very common source for wide range of human activities, for all we can mention MR, CT, SPECT images in medicine or NIR (near infra red) and UV images in restorers workouts. Multimodality of these resources is temporal as well as based on technology used for capturing source images. Processing of multimodal images is composed mainly of image comparison (i.e.: pre and post images of surgery operative or spectroscopy of materials). In many cases image comparison is not trivial task and can be made only by specialists. Our goal is to facilitate the work of experts by highlighting parts of image which ”differ” more than other (regions which carry more relevant information than other). This intention is poorly specified problem due to various requirements of each comparison. So we resign for any robust method giving crucial results but we propose method providing better overview on data sets. Previous work In spite of wide usability of difference metrics image processing field is very poor in algorithms offering difference measures. An approach to this problematics we can see in [Ulstadt, 1973] or in [der Weken et al., 2004] where the difference is computed over normalized images. And in many other papers we could pass metrics based on MMSE (minimal mean square error) i.e. [Moon and Kim, 1994], but these metrics are unusable for multimodal data because we cannot assume Gaussian distribution of colours neither any distribution equal for both input images. Therefore central moments are for our purposes unusable. Different modalities are used for extra features which are mostly unpredictable from any other modality. These features are often two colours separable in only one modality, details in one modality versus colours in second, changed objects in multi-temporal images. By this reason we can also reject all approximation approaches in tasks where number of scene classes is unknown in any modality. Our approach goes out from image registration [Viola, 1995] and image fusion algorithms mainly because these algorithms operate with multimodal data more sensitively than common difference metrics. However algorithms for image registration solve different problems (looking for referring points) and image fusion algorithms often combines multimodal data by different characteristics (i.e.: details and colours, segments, etc.) in each image. Usage of multimodal data sets In the figure 2 we can see joint histogram of two multimodal images. For a sketchy view to multimodal data we mention two peaks at point [240, 62] and [240, 162]. These points represent two colours in NIR spectrum but only one colour in visible spectrum. The first goal is highlighting of regions which correspond with two different colors in NIR spectrum but in visible spectrum are inseparable. These different NIR regions we dye by green and yellow colour see figure 2b. As can be seen on the dyed image, this comparison disregards any structure of the image, just separates pixels from different NIR classes. Instead of separating regions (this is task for image fusion), we should look into structure of the image (edges, segments, object classes) and highlight these parts of the image, where classification differs. 37 WDS'10 Proceedings of Contributed Papers, Part I, 37–41, 2010. ISBN 978-80-7378-139-2 © MATFYZPRESS