Taking the Most of Existing Image Descriptors: A Hybrid Approach

摘要

This work proposes the creation of a new image descriptor based on the aggregation of multiple pre-existing image descriptors. It starts by generating intermediary descriptors based on existing extraction procedures. Afterwards, those descriptors are normalized and eight first order statistical measurements are computed for each individual descriptor. It follows a step of parameter selection which aims to remove biased and irrelevant features. At the same time, it seeks to select the features which maximize the amount of information encoded. The resulting descriptor was tested using six different image databases, namely two produced in-house and four well-known public bases UIUCTex, Brodatz, VisTex and Outex TC 00013. The results seem to indicate that superior descriptive power can be achieved using this selection procedure.