Case Studies

Case Studies

Raoul Pascal Pein (University of Huddersfield, UK), Joan Lu (University of Huddersfield, UK) and Wolfgang Renz (Hamburg University of Applied Sciences, Germany)
Copyright: © 2013 |Pages: 41
DOI: 10.4018/978-1-4666-1975-3.ch024

Abstract

This chapter discusses several case studies to evaluate the methods introduced in chapter 23, Methods. Each case study focuses on a specific issue and the advanced cases build up on previous findings. The first ones are dealing with the low-level fv directly and then the scope widens to the interrelationship of multiple fv until their combination within a single query is used as a mapping rule for higher-level semantics (i.e. categories).
Chapter Preview
Top

Feature Normalization

As mentioned in section Methods-Feature Normalization, each fv creates a different similarity curve between 100% match and the least relevant document. Similarities of features are not comparable to other features. For that reason, it is attempted to capture the fv specific similarity profiles and use them to normalize the similarity values within each retrieval.

Requirements

  • •

    Default image database with use-case related content (e.g. photographs)

Testing

First, the original similarity profile of each feature vector plug-in is determined for both image repositories (ETH-80 and Caltech-101). For each ranking position, the average, minimum and maximum value are collected. These profiles are used as reference to create the normalization functions and to measure the changes according to the normalized profiles. Each average curve is split into 20 segments of equal size. The selected boundaries are 1.0 for rank0 and 0.0 for rank|I|+1. In between, the average similarity of each of the 19 equidistant ranking positions are used as reference points for the later normalization. The feature vector modules used: RGB Mean, Histogram, Spatial Histogram, Wavelet

The normalized similarity profiles are expected to be close to the desired function in section Methods-Features and Similarity Measures-Feature Normalization-Determining a Normalization Function. Especially for the dataset specific normalizations, the average values should produce only a small error. Normalization parameters extracted from a different image set are expected to result at least in a better profile than without normalization.

Complete Chapter List

Search this Book:
Reset