UNIVERSITY OF SOUTHAMPTON COMP6223W1 SEMESTER 1 EXAMINATION 2016 – 2017 COMPUTER VISION (MSC) DURATION 120 MINS (2 Hours) This paper contains 6 questions Answer THREE questions. An outline marking scheme is shown in brackets to the right of each ques- tion. University approved calculators MAY be used. A foreign language dictionary is permitted ONLY IF it is a paper version of a direct Word to Word translation dictionary AND it contains no notes, additions or annotations. 7 page examination paper. Copyright 2017 c University of Southampton Page 1 of 7 COMP6223W1 Question 1. (a) Explain what is meant by edge detection in computer vision. De- scribe the difference in principle between first- and second-order edge detection. [9 marks] (b) Provide a pseudocode description of the Laplacian operator which, given a grey-level image as input, delivers a binary image where points are ‘1’ where an edge occurs and ‘0’ otherwise. Explain pre- cisely how your code should operate, and justify any choices you have made in your implementation. [16 marks] (c) Describe two ways in which the basic Laplacian operator can be made less sensitive to noise. Discuss the relative advantages and disadvantages of your approaches. [8 marks] Copyright 2017 c University of Southampton Page 2 of 7 COMP6223W1 Question 2. (a) Describe the aims and differences between the processes of inten- sity normalisation and histogram equalisation. [9 marks] (b) A monochrome camera is known to have a poor response to low light and an excessive response to bright illumination. For grey levels between zero and 127 the gain is 0.5 whereas for grey levels between 128 and 255 the gain is 1.5. Sketch the relationship between camera output and grey level. Describe using pseudocode an operator that normalises the output so that the effective camera gain for all grey levels is 1.0. [16 marks] (c) One (rather dated) approach to find image features of interest is to apply histogram equalisation followed by optimised thresholding. Dis- cuss advantages and limitations of your new approach developed in part (b), followed by a form of thresholding, in comparison with the histogram equalisation based approach. [8 marks] Copyright 2017 c University of Southampton TURN OVER Page 3 of 7 COMP6223W1 Question 3. (a) Show the bases of the Hough transform for conic sections wherein the Cartesian parameterisations of a line and of a circle can be viewed as a parameter space analysis. [10 marks] (b) Show how the foot-of-normal parameterisation of a line is derived. [14 marks] (c) Describe how the foot-of-normal line parameterisation limits the pa- rameter ranges leading to a practical implementation of the HT for lines. [9 marks] Copyright 2017 c University of Southampton Page 4 of 7 COMP6223W1 Question 4. (a) State the two categories in which shapes (represented by connected components) can be described. For each category, provide details of a specific descriptor, and briefly describe how it can be computed from a connected component. [8 marks] (b) Consider the connected component depicted by the solid black pixels below: Ensuring you show all working, compute: • the compactness • a 4-connected chain-code representation [11 marks] (c) Describe in detail the process of creating a Point Distribution Model to describe the shape of a human face. [14 marks] Copyright 2017 c University of Southampton TURN OVER Page 5 of 7 COMP6223W1 Question 5. The government of the country of Taghum wants to introduce alpha- numeric license plates for the owners of motorised vehicles, and roll-out a state-wide surveillance programme to track vehicles using Automatic Number Plate Recognition (ANPR) using Computer Vision. You have been asked to design the system. The government has specified that it is expecting its license plates to be around the same size as those used in countries like the UK, and to use some combination of letters from the English alphabet and digits from standard (Arabic) numerals. (a) Given that vehicles in Taghum do not currently have license plates, describe how you would design the licence plates and hardware as- pects of the computer vision system for performing ANPR. State the rationale for your design choices. [15 marks] (b) Starting with an image captured by your ANPR hardware, describe in detail the processing that your system will perform to recognise the individual characters within a licence plate. [18 marks] Copyright 2017 c University of Southampton Page 6 of 7 COMP6223W1 Question 6. (a) Describe in detail how the Harris and Stephens Corner Detection algorithm works. [15 marks] (b) Given the following Structure Tensor, 1600.0 50.0 50.0 1600.0 assuming k = 0.04, compute the Harris Corner Response, showing all working. Sketch the type of image patch that this structure tensor is likely to belong to. [8 marks] (c) Briefly describe how SIFT features are computed and give details of a robust method for finding correspondences between two images using interest points described by SIFT features. [10 marks] Copyright 2017 c University of Southampton END OF PAPER Page 7 of 7
