The assignment is to implement the above user’s story (To develop and implement Image Denoising using Generative Adversarial Networks which you will use Least Squared Methods as the loss function instead of Wassertein-GAN as the loss function) with Python programming Language and compared the result with theirs “(A generative adversarial network for image denoising
Yue Zhong1,2 · Lizhuang Liu2 · Dan Zhao2 · Hongyang Li2
Received: 6 December 2018 / Revised: 30 January 2019 / Accepted: 28 March 2019 /© Springer Science+Business Media, LLC, part of Springer Nature 2019)” results using PSRN as the performance metrics.
1.1 Statement to the problem Traditional image denoising algorithms such as Linear Filter, Min Filter, Median Filter, Wiener Filter, and others presume that the noise is uniformly distributed and homogeneous [17]. In practice, however, noise in real-world images like high-ISO pictures and microscopic fluorescence images is more complicated [3]. [16] Real-noise or Blind-noise is the term for such noise on real pictures. Since the acquisition of such real images is difficult, conventional filters struggle to work well on images with such noise [18].[16] proposed “A generative adversarial network for image denoising” using DenseNet framework as a generator and Wasserstein distance as loss function which produce a photo-realistic image with higher quality. Despite this cutting- edge technique, it still fails to fully separate noisy pixels, particularly in the presence of unknown noise because of the problems associated with the WGAN such as Vanishing Gradients, which occurs when the discriminator becomes too successful and the generating gradient vanishes, causing the discriminator to learn nothing, the model parameters oscillate, destabilize, and never converge due to a non-convergence problem. Moderate Collapse allows the generator to collapse, resulting in a limited variety of samples, Overfitting is caused by an imbalance between the generator and the discriminator and are quite sensitive to the choice of hyperparameters, However, the noisy photos are denoised to some extent, but this results in another noisy image [16]. 1.2 Motivation Different intrinsic (sensor) and extrinsic (environment) conditions can cause image noise, which is often impossible to prevent in practice [19]. As a result, image denoising is useful in a variety of applications, including image preservation, visual monitoring, image registration, image segmentation, and image classification, where acquiring the original image content is critical for good results [20]. Although several algorithms have been proposed for image denoising, the problem of image noise suppression remains unsolved, especially in situations where images are acquired in poor conditions with a high level of noise [21][16]. However, the proposed study aims to develop an architectural model with Generative Adversarial Models for image denoising that can denoise an unknown noise using the least square method as the loss function [22] 1.3 Aim and Objectives This research aims to create and test an architectural model for image denoising that uses the Least Squared Generative Adversarial Network (LSGAN) model. The following are the goals of this study: i. Create an improved network architecture for image denoising that uses the least-squares method as the loss function. ii. Implement the enhanced network architecture for image denoising iii. Using PSRN, compare the output of the previous models. 1.4 Network Architecture of Existing and Proposed Model Figure 1.4: Network Architecture of Existing system Figure 1.5: Network Architecture of Proposed system ASSIGNMENT TO YOU The assignment is to implement the above user’s story (To develop and implement Image Denoising using Generative Adversarial Networks which you will use Least Squared Methods as the loss function instead of Wassertein-GAN as the loss function) with Python programming Language and compared the result with theirs “(A generative adversarial network for image denoising Yue Zhong1,2 · Lizhuang Liu2 · Dan Zhao2 · Hongyang Li2 Received: 6 December 2018 / Revised: 30 January 2019 / Accepted: 28 March 2019 /© Springer Science+Business Media, LLC, part of Springer Nature 2019)” results using PSRN as the performance metrics.