AbstractContour detection is the process of finding meaningful smooth curves within an image， such as the boundaries of objects. It is a process at which the biological visual system excels and is also a fundamental technique and crucial step for many applications in computer vision. This thesis is primarily concerned with biologically-inspired approaches to contour detection. In this thesis, some biologically-inspired contour detection algorithms were improved. Such improvement included the better performance obtained from different extended versions of an existing biologically-inspired method, simplification of a non-biologically-inspired method to a biologically-inspired method without degrading performance, and application of a biologically-inspired method to improve other biologically-inspired and non-biologically-inspired methods. All of the contour detection results mentioned in this thesis were both qualitatively and quantitatively evaluated by the BSDS300 and/or BSDS500 benchmarks using the quantitative measures: F-score and/or ODS, OIS and AP. All the parameter values of the evaluated methods were determined by operating the algorithm on training images in this benchmark and observing the alteration of F-score or ODS.
A neurophysiologically-inspired model, the PC/BC model of V1 devised to locate boundaries defined by intensity discontinuities, was first extended to also be able to locate boundaries defined by colour discontinuities. This extension was inspired by neurophysiological data from single neurons in macaque V1. The colour PC/BC model of V1 (F-score 0.67) have significantly better performance compared with the original PC/BC model of V1 (F-score 0.61) and slightly outperform some recently proposed contour detection algorithms which use more cues and/or require a complicated training procedure. The PC/BC model of V1 was also extended to simulate the second visual cortical area (V2) in an attempt to detect texture boundaries and corners. The two versions of the PC/BC model of V2 for texture boundary and corner detection were integrated with the PC/BC model of V1 to successfully locate some texture boundaries and enhance some weak contours obtained from the PC/BC model of V1. However, the integrated PC/BC model for texture boundary detection (F-score 0.60) and corner detection (F-score 0.51) both generated worse performance compared to the original PC/BC model of V1 (F-score 0.61).
A simplified and biologically-inspired version of the texture gradient method used in the probability of boundary (Pb) algorithm is also proposed in this thesis. The Pb algorithm is the first stage of a recent near state-of-the-art contour detection algorithm. The proposed texture gradient method (F-score 0.58) has similar performance in detecting texture boundaries to the Pb texture gradient method (F-score 0.58) but also reduces computation time by about 240 seconds when tested on 100 images and simulates texture boundary detection in V2 of human/animal visual system. Finally, in this thesis, a modified version of the sparse reconstruction-based discrimination (SRBD) method inspired by the neurophysiological evidence for sparse coding in the biological visual system was applied to refine the contour detection results produced by the colour PC/BC model of V1 and two leading contour detection algorithms. Previously, the SRBD method (F-score 0.66) had been applied to significantly improving the contour detection results of the Canny edge detector (F-score 0.58). Here, the modified version of the biologically-inspired SRBD is shown to improve the result derived from three previously proposed contour detection methods with at least one of the quantitative measures used in the BSDS500 benchmark increased by 0.01.
|Date of Award||2017|
|Supervisor||Michael Spratling (Supervisor) & Thrishantha Nanayakkara (Supervisor)|