Convolutional Neutral Network Of Bilateral Filter Image Processing

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.