PROJECT TITLE :

Fast Adaptive Bilateral Filtering

ABSTRACT:

For edge-preserving smoothing, a fixed Gaussian range kernel and a spatial kernel are employed in the bilateral filter. It is possible to generalise this filter by allowing the centre and breadth of the Gaussian range kernel to change from one pixel to the next. In addition to sharpening and noise removal, this variation can also be used for other purposes, such as artefact removal and texture filtering. The brute-force implementation of its adaptive cousin, like the bilateral filter, involves a lot of computations. However, despite the existence of numerous fast bilateral filtering methods in the literature, most of them only function with a fixed range kernel. A fast approach for adaptive bilateral filtering that does not scale with the spatial filter width has been proposed in this study. That's because range-based filtering may be done using a suitably specified local histogram in this case. Analytic functions can be used to approximate the filter by replacing the histogram with a polynomial and the finite range-space sum with an integral. Fast convolutions can be used to match polynomial moments to target histogram moments, and the analytic functions are iteratively computed using integration-by-parts to achieve an efficient approach. The brute-force implementation can be sped up by at least 20 with our technique without affecting the visual quality. Sharpening, JPEG deblocking, and texture filtration are all shown to be effective using our technique.