A new restricted boltzmann machine training algorithm for image restoration
- PDF / 1,654,329 Bytes
- 16 Pages / 439.642 x 666.49 pts Page_size
- 2 Downloads / 228 Views
A new restricted boltzmann machine training algorithm for image restoration Ali Fakhari1 · Kourosh Kiani1 Received: 3 May 2019 / Revised: 14 August 2020 / Accepted: 20 August 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract A variety of approaches have been proposed for addressing different image restoration challenges. Recently, deep generative models were one of the mostly used ones. In this paper, a new Restricted Boltzmann Machines (RBM) training algorithm for addressing corrupted data has been proposed. RBMs can be trained both supervised and unsupervised, however they are very sensitive to noise and occlusion. The proposed algorithm enables the RBM to be robust against corruptions. Using the new algorithm, we have given the RBM a posterior knowledge about desired or clean data. Despite other methods, the proposed algorithm works fine without changing the architecture of the model or adding any regularization term. Concretely, the RBM can be used as a robust feature extractor, even for unclean data. By creating different corrupted versions for each image instance, and using the original version in the reconstruction phase, the RBM can learn the desired probability distribution of data. Experimental results confirm the robustness of the model against different types of corruption. Keywords Image restoration · Generative models · RBM · Contrastive divergence
1 Introduction Image restoration problem can be defined in different categories, including image denoising, inpainting, deblurring, super resolution. The common thing among all mentioned problems is that the restoration task has to remove the corruption (H ) from the images, regardless of the corruption type. The problem is that the corruption function is usually unknown. Considering the joint probability between corrupted input and desired outputs, the image restoration task can be done without any information about corruption function. As generative models consider the joint probability between their input and output, they can be Kourosh Kiani
[email protected] Ali Fakhari [email protected] 1
Electrical and Computer Engineering Faculty, Semnan University, Semnan, Iran
Multimedia Tools and Applications
employed as a good choice for addressing this kind of problems. (1) demonstrates the image restoration task using the joint probabilities of the original and the corrupted images. P (R, C) = P (R) × P (C|R)
(1)
where R is the original image and C is the corrupted image distorted by H . As can be seen in this formulation, i.e. joint probability of original and corrupted images, there is no need to know the H or corruption function. As in most image restoration problems H is unknown. Considering the joint probability of the input and the output, Generative models learn the true data distribution so that they can generate new data samples according to the learned distribution. RBM is one of the most frequently used deep generative models employed in different applications. RBMs are undirected probabi
Data Loading...
