A Faster than Real-Time Implementation of Multi-exposure Image Fusion with Recursive Filter

This paper proposes a faster than real-time GPU implementation of multi-exposure image fusion with a recursive filter. The algorithm transfers the images from RGB space to YUV space first, then weight fuses them in Y, U, V channels separately with recursive filters, at last transfers the combined image to RGB space again and get the final output fusion image. We further propose novel task-level parallelism, parallel dataflow reunion and parallel recursive filter techniques to accelerate the image fusion algorithm on the GPU platform. In our experiment, we obtain 15-20 times speedup compared to the CPU version and up to 100 fps in a 600*600*6 multi-exposure sequence image fusion case. Experimental results demonstrate the superiority of our proposed method in terms of high efficiency and high availability.