Use of Weighted Visual Terms and Machine Learning Techniques for Image Content Recognition relying on MPEG-7 Visual Descriptors

摘要

We propose a technique for automatic recognition of content in images. Our technique uses machine learning methods to build classifiers which are able to decide about the presence of semantic concepts in images. Our classifiers exploit a representation of images in terms of vectors of visual terms. A visual term represents a set of visually similar regions that can be found in images. Various types of visual terms are used at the same time to take into account various similarity criteria and region representations that are available to compare regions. Specifically, we compare regions using the 5 MPEG-7 visual descriptors. An image is indexed by first using a segmentation algorithm to extract its regions, and then the image is associated with the visual terms that are more similar to the extracted regions. The proposed technique offers very good performance as demonstrated by the experiments that we performed.