Binary image comparison based on a local measure of dissimilarity.
Application to classification
Following the beginning of the national program of ancient document digitalization, the Troyes' library, which has such documents has decided to participate to it. This has led to a collaboration between the library and the image team of the laboratory CReSTIC so as to develop the methods and computer tools to exploit the resulting database.
This PhD deals with the comparison of binary images that are not composed of a single pattern. The proposed method is then extended to gray level images. Using a measure example - the Hausdorff distance (HD) - a local measure is defined throught a window, and its properties depending on the window size and the global HD measure are proved. Thanks to them, a window-size criterion is defined so as the window to be adjusted to the local dissimilarity. The local dissimilarity map (LDM) is then defined when the window slides over all the compared images. The LDM can be defined with other measure than the HD using the same algorithm, nevertheless, the DH properties leads to a LDM fast computation. The LDM can be used as image dissimilarity visualization method or a tool to decide on image similarity. For this last point, a first step is a binary-image adapted multiresolution analysis which is base on the median morphological filter. This allows to have an resolution adapted to the researched similarity degree. A second step consists in using LDM information concerning the dissimilarities and their spatial distribution to compare the images. Several comparison methods are tested, the most efficient one is based on the SVM with the whole LDM as input data. The method efficiency is successfully tested on an ancient-impressions database and on a face database.
BaseThe PhD method has been tested on an experimental database.
Experimental base of medieval impression (68 items) : base.zip
The study has given a method of local comparison. It allows to compare the binary images that are not composed of a single pattern and then to classify them. there are two ways to explore:
Other project about the patrimonyMaDonne