Dr. Alasdair D. F. Clarke<\/h1>\n

I am now working at Informatics, University of Edinburgh. My old email still works for now, and I’ll put a link to my new homepage when I get round to making it!<\/p>\n

The Texture Lab,
\nSchool of Mathematical and Computer Science,
\nHeriot-Watt University, Edinburgh, EH14 4AS
\nadfc1 at hw dot ac dot uk
\n+44 (0)131 451 4166<\/p>\n

Publications<\/a>
\nBiography<\/a>
\nPhD Thesis<\/a><\/p>\n

Chapter 1: Introduction<\/h3>\n

Automatic surface defect detection is one of the main applications of computer vision and many different approaches and methods have been put forward. However, the ability of the human vision system to detect surface defects has not been studied in a rigorous way and little effort has been made to investigate how well computer vision algorithms can mimic human behaviour. Hence the aim of this thesis is to bring together relevant work on visual search, saliency, perception and texture discrimination for the purpose of analysing modelling human defect detection.<\/p>\n

Chapter 2: Lit Review<\/h3>\n

This chapter contains two literature reviews. Section 2.1 concerns computer vision and reviews some of the different approaches that have been used to tackle the problem of automated surface defect detection. The related problem of texture discrimination is also discussed. The second half of this chapter, Section 2.2, contains a general overview of the processes behind human perception and introduces the \u001cfield
\nof visual search.<\/p>\n

Chapter 3: Texture Synthesis<\/h3>\n

Chapter 4: Visual Search for a Defect on a Homogeneous Surface<\/h3>\n

In this chapter I will explore how human observers perform in a search task involving a target on a homogeneous surface texture. I will investigate how surface and target properties affect our ability to \u001cfind a defect in a forced choice target absent\/present task using the two different surface textures described in the previous chapter. For the 1\/f\f-noise surface textures the effect of varying surface roughness, along with the depth and orientation of the target, will be explored, while for the near-regular surfaces I will vary texton density and the degree of regularity. Altogether, four experiments will be carried out. The use of an eye-tracker will allow for search strategies to be investigated and a computational saliency model [Itti and Koch, 2000,Walther and Koch, 2006] will be run on the experimental stimuli and compared to the results from the psychophysical experiment.<\/p>\n

This chapter starts with a discussion of the concept of visual saliency, computed by bottom-up visual processes (Section 4.1). This review is centred around Itti and Koch’s [2000] saliency model and the extent to which it can explain visual search paths. This is followed by a methods section in which the procedures for the experiments in this, and Chapter 6, are given (Section 4.2). The core of this chapter contains a series of four experiments designed to investigate how well human observers can carry out a visual search task on the textural stimuli discussed in the preceding chapter (Section 4.3). Finally, results from the human observers are compared to the performance of the saliency model in Section 4.4.<\/p>\n

Chapter 5: Models of Visual Search<\/h3>\n

In the previous chapter I compared Itti and Koch’s visual saliency algorithm to human performance in a series of visual search experiments involving a defect on an otherwise homogeneous textured surface. While the model proved a good match for the human data when a circular indent was used, it failed to mimic human behaviour when an elongated target was used. In this short chapter I will give a comprehensive review of previous work on modelling visual search before designing my own model in Chapter 6.<\/p>\n

Chapter 6: An LNL-Based Search Model<\/h3>\n

In Chapter 4 I investigated how human performance in a defect detection task varies with surface and target properties such as regularity and orientation. Human performance was also compared to a bottom-up visual saliency algorithm [Itti and Koch, 2000, Walther and Koch, 2006]. The results showed that the model only partially \u001ctted the human data: in particular there is a discrepancy between the performance of the model and human observers when searching for an elongated defect (Experiment 3).<\/p>\n

Chapter 5 contained a comprehensive review of how the problem of modelling visual search has been tackled in the past and in this chapter I will attempt to construct a search model which can simulate human behaviour in an unsupervised surface defect detection task. The model is based on an LNL-framework (see Section 2.1.2) and will be compared to human performance in a series of experiments, using a target always present design. This will remove the need for considering speed\/accuracy trade-offs. Furthermore, it avoids the problem of de\u001cning a decision rule for the model for target absent trials. Instead, the model is assumed to \u001cnd the defect when it \u001cxates on it and only one measure, the number of saccades needed to \u001cfind the target, needs to be used in order to compare performance between the search model and the human observers.<\/p>\n

Chapter 7: Stochastic Search Strategies<\/h3>\n

In the previous chapter a LNL-based search model was shown to provide a good prediction of the \u001edifficulty of fi\u001cnding a defect in a rough surface. The model was tested over a wide range of perceived surface roughnesses and task difficulties and made a similar number of saccades to human observers in all cases. However, the model did a poor job of accounting for human scan-paths and search strategies and only predicted human \u001cfixation locations at chance levels. This suggests that signal-to-noise ratio in the activation map generated by the LNL-based model is a good model for human performance, but the local maxima in the activation map do not provide a good predication of fi\u001cxation locations.<\/p>\n

In this chapter I will explore search strategies and how much of a role visual memory has in determining search performance. The \u001cfirst experiment (Section 7.3) will investigate memory using a moving target paradigm [Horowitz andWolfe, 1998]. This is followed by a comparison between human performance and a stochastic search simulation. Unlike the previous model which was concerned with feature extraction, the stochastic search simulation only attempts to model search strategy and saccade choice. The model is outlined in Section 7.4 and uses the results of a signal detection experiment (Section 7.5) for the target detection model. The model is compared with human observers in Section 7.6.<\/p>\n