Feature selection and extraction of visual search strategies with eye tracking.
File(s)
Author(s)
Hu, Xiao-Peng
Type
Thesis
Abstract
Visual attention is a selective activity of visual search. It stems from the limited ability of our human visual system, and plays an important role in many aspects of visual processing including feature detection, pattern recognition, object identification, and reasoning. The performance of visual search depends heavily on the strategy of deploying visual attention in response to different search tasks. The need to understand visual search strategies has been recognised in both theoretical and practical vision research for many years. Although it has been well established that visual search can be divided into pre- attentive and attentive stages, deterministic models that can predict fixation positions and transitions are not yet available. To date, there has been limited work for identifying the reasons for the choice of fixation points. The method by which observers glean certain subtle features and combine this information with other image features remains poorly understood. The aim of this research is to investigate the theoretical and practical issues involved in visual search strategies for attention selection and knowledge gathering. A new framework for hot spot detection is proposed for capturing the intrinsic visual search behaviour of different observers in image understanding. The method is based on information theory for identifying salient image features used for visual search. We demonstrate how to obtain feature space fixation density functions that are normalised to the image content along the scan paths. This allows the reliable identification of salient image features that can be mapped back to spatial domain for highlighting regions of interest and attention selection. To fully understand visual attention, we also propose an algorithm for extracting the pattern of fixation sequences, which can be crucial for object recognition in order to establish the relationship between visual features and resolve ambiguities. We present a feature selection algorithm based on Bayesian theory for enhancing the understanding of the interaction between top-down and bottom-up processes in human visual perception. We demonstrate the relationship between filter- based feature selection and the performance of ideal inference, in which the expected AUC (the Area under Receiver Operating Characteristic Curve) is employed to describe the intrinsic discriminability of features. A computationally efficient algorithm is designed to reduce the complexity of the feature selection process. With this thesis, we will also propose a Bayesian method for the integration of explicit domain knowledge for scan path analysis, thereby providing a theoretical basis for modelling visual feature selection. The proposed framework is validated with both artificial and real-world examples, highlighting the robustness and strength of the techniques. The practical value of the method for knowledge gathering for diagnostic decision support in medical imaging is demonstrated with high-resolution Computed Tomography lung image analysis.
Version
Open Access
Date Awarded
2005
Copyright Statement
Attribution NoDerivatives 4.0 International Licence (CC BY-ND)
Advisor
Yang, Guang-Zhong
Sponsor
Engineering and Physical Sciences Research Council (EPSRC): Royal Society/Wolfson Medical Image Computing Laboratory.
Grant Number
Visual Tracking for Active Learning - ViTAL (GR/N08810).
Publisher Department
Department of Computing.
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)
Author Permission
Permission not granted