| dc.description.abstract | Gaussian probability distribution functions summarize the distribution of random variables, and the characteristics of the functions, such as the parameters of the
procedures, are outlined by Gaussian processes. As a result, Gaussian processes may be
considered one degree of abstraction or indirection above Gaussian functions. For predictive classification modeling, Gaussian processes can be employed as a machine learning method. Gaussian processes, like support vector machines, are a form of kernel
technique; however, unlike support vector machines, they can predict highly calibrated
probabilities. In image processing, the Gaussian model is also essential in terms of image
quality improvement and modeling the constraints of pixels in the spatial domain.
A bilateral filter is a nonlinear picture smoothing filter that preserves edges
while lowering noise. It uses a weighted average of intensity data from surrounding
pixels to replace the intensity of each pixel. A Gaussian distribution can be used for
this weight. Notably, the weights are based on the radiometric discrepancies and the
Euclidean distance between pixels (e.g., range differences, such as color intensity, depth
distance, etc.). Sharp edges are preserved as a result. Bilateral filtering is based on the
notion of doing what typical filters do in their domain in the range of an image. Two
pixels can be adjacent to one another in the sense that they occupy neighboring spatial
locations. They can be similar because they have comparable values, potentially in a
perceptually meaningful way. However, the bilateral filter depends on a Gaussian distribution that borrows terms from statistics; it is challenging to implement a bilateral
filter with real-time computing. When executing the traditional implementation
method and bilateral functions in common libraries such as OpenCV, Skimage, we only
use CPUs. At the same time, modern computers are all equipped with GPUs, which
means we are wasting existing resources because of not using contemporary computer
pieces of equipment.
In this thesis, we proposed a bilateral filtering approach implemented in the
form of a convolutional neural network and parallel computing technique. Our method15
has been experimented with the Color BSD68 dataset and National Geographic's "Sublime Scenery" video. Our approach achieved a lower peak signal-to-noise ratio in comparison with OpenCV's bilateral filtering methods. In terms of performance, our
method reached an average processing speed approximately at 107 FPS with Color
BSD68 dataset and approximately at 127 FPS with the chosen video, which means
that our approach is well-suited for real-time application. | en_US |
