GPU-Accelerated Light-Field Image Super-Resolution

Light-field imaging has become an emerging technology that brings great benefits to many fields, i.e. in photography, academia, and industry. However, these benefits come with the cost of high computation requirement that limits its applications in practice. This paper presents an accelerated solution for 4D light-field image super-resolution. The acceleration is achieved by the mean of parallel computation using graphics processing units. The selected algorithm is broken into functions which is suitable for parallel execution. Each of the functions is then transformed into GPU kernel and executed at each work-item which is associated with a pixel location in the proposed architecture. Using disparity maps extracted from input 4D light-field as an aid for super-resolution task, the proposed approach can successfully super-resolute an input 4D light-field by the factor of 4 horizontally and vertically. Two strategies, Y-RGB and RGB, are proposed to handle color images. Y-RGB is suitable for high-speed processing constraints while RGB is more preferable if output quality is the main concern. Experimental results show that the proposed approach can achieve the speed up of 203× and 71× compared to CPU implementation for Y-RGB and RGB strategy respectively. Regarding output quality, the proposed approach generates a shaper high-resolution image with more details compared to the baseline methods.