The ubiquity of camera and display devices provides a unique opportunity for low-cost realizations of widespread device-to-device communications in the specific form of display-camera visible light communications. A major impairment of the display-camera communication channel is the perspective distortion, which is a non-linear transform acting on the pixel locations of a displayed image/video. Conventionally, dedicated training patterns are required for estimating the perspective distortion parameters, which determine how the locations of the image/video captured by a camera are distorted. In this thesis, a novel algorithm of perspective distortion estimation for images captured from a display device without requiring training patterns is presented. It takes advantage of the p...[ Read more ]

The ubiquity of camera and display devices provides a unique opportunity for low-cost realizations of widespread device-to-device communications in the specific form of display-camera visible light communications. A major impairment of the display-camera communication channel is the perspective distortion, which is a non-linear transform acting on the pixel locations of a displayed image/video. Conventionally, dedicated training patterns are required for estimating the perspective distortion parameters, which determine how the locations of the image/video captured by a camera are distorted. In this thesis, a novel algorithm of perspective distortion estimation for images captured from a display device without requiring training patterns is presented. It takes advantage of the pixel grid structure of common display devices, which can be observed as the screen door patterns of the captured images, as an implicit training pattern.

A geometric model of the screen door pattern under perspective distortion is proposed and its validity has been verified using captured images from common display devices. Based on the proposed geometric model, the problem of perspective distortion estimation is formulated as an optimization problem. The display pixel grid lines associated with a captured image are detected using the modified Hough transform and are used to obtain coarse estimates of the three angle parameters of the perspective distortion. Together with some known physical parameters of camera devices, these three angles are used to construct an initial guess solution. An iterative algorithm is derived to produce a much improved solution to the formulated optimization problem in estimating the modified perspective distortion parameters. Using the estimated perspective distortion parameters, perspective correction is applied to the captured image to reconstruct the original image. Since both of the displayed and captured images are digital with signal samples in discrete pixel locations only, natural interpolation is used in the perspective correction process. Experimental results show that satisfactory estimation accuracy for perspective correction can be achieved by the proposed algorithm for images with various types of content.