A Comprehensive Analysis Of Straight Line Estimation With A Novel Noise Dataset

Accurately detecting straight lines in an image within a short time frame is a fairly easy process for a human being, while A.I. systems still struggle despite recent advancements in the field. This challenge has seen various limitations across the decades from the need for powerful hardware to the lack of training data or unreliable detectors. This thesis presents a novel noise dataset called Artificially Generated Objects with Noise (AGON), with the goal of advancing the state-of-the art in straight line segment detection and noise corruption research. Generating 27,195 synthetic images under 11 noise model distributions and 36 object rotation angles, AGON represents the baseline line annotated dataset for a detailed performance analysis of artificial noise models available in digital cameras. Exploring the effects of noise on eight traditional or new line detectors and their parameter dependency, this thesis demonstrates that parameterless methods achieve better results, only outputting true positives. Introducing round edge shapes in this dataset outlined the wider scope of the straight line detectors of reliably detecting all types of object outlines. Using 8 methods with at least four relatively distinct ways of detecting a line, for all techniques, the result was similar: the more noise is present in a 2D image, whether synthetic or real, the greater the loss of information and the more time each segment validation takes.