Project: Evaluation of Visual Concept Classifiers

Team: Christian Hentschel, Dr. Harald Sack

Research institution: Hasso-Plattner-Institut Potsdam

Abstract: Approaches for content-based visual concept classification heavily rely on extraction of local and global visual features such as color and gradient histograms represented in high-dimensional vector spaces. Appropriate machine learning algorithms can be applied to separate positive from negative examples for a specific visual concept (i.e. contents category). The very successful Bag-of-Visual-Words (BoW) approach for visual content classification follows an idea borrowed from text retrieval domain: An image is described by a frequency distribution of visual words represented by densely sampled local histograms of gradients. Support Vector Machines (SVM) are applied to separate positive from negative examples of a specific contents category by evaluating the similarity between different training examples. While this approach provides promising results a major drawback, however, is the impossibility to infer the impact of the various visual words, i.e. the various dimensions of the image descriptors, on the overall classification result. This information is important when trying to improve the overall classification accuracy by improving the process of selecting meaningful visual words. Subject of the proposed research project, thus, is the implementation and evaluation of methods to gain these information through analysis of the obtained SVM models and application of alternative machine learning approaches. Visualization of the impact of a visual word for a specific category classification will be done through heat maps.

Last modified 8 years ago Last modified on Apr 13, 2013 10:59:43 AM