Fully Convolutional Pixel Adaptive Image Denoiser

Abstract

We propose a new image denoising algorithm, dubbed as Fully Convolutional Adaptive Image DEnoiser (FC-AIDE), that can learn from offline supervised training set with a fully convolutional neural network as well as adaptively fine-tune the supervised model for each given noisy image. We significantly extend the framework of the recently proposed Neural AIDE, which formulates the denoiser to be context-based pixelwise mappings and utilizes the unbiased estimator of MSE for such denoisers. The two main contributions we make are; 1) implementing a novel fully convolutional architecture that boosts the base supervised model, and 2) introducing regularization methods for the adaptive fine-tuning such that a stronger and more robust adaptivity can be attained. As a result, FC-AIDE is shown to possess many desirable features; it outperforms the recent CNNbased state-of-the-art denoisers on all of the benchmark datasets we tested, and gets particularly strong for various challenging scenarios, e.g., with mismatched image/noise characteristics or with scarce supervised training data.

DoPAMINE: Double-Sided Masked CNN for Pixel Adaptive Multiplicative Noise Despeckling

“…Figure 3(e) compared the PSNR of F T and AF T on DoPAMINE. We see the maximum PSNR of AF T is 0.3dB higher than F T , and this is in line with the findings in (Cha and Moon 2018a). This result also show that the performance of F T and AF T is robust to epochs around 5 to 15.…”

“…To overcome such limitation, (Cha and Moon 2018b) have recently proposed an adaptive method, dubbed as N-AIDE, for the additive noise case that can carry out both the supervised training (on an offline dataset) and adaptive fine-tuning (on a given noisy data) of the network parameters. Later, N-AIDE was extended in (Cha and Moon 2018a) to implement a fully convolutional architecture and unknown noise variance estimation scheme. The crux of N-AIDE is to design the neural network to learn pixelwise affine mappings with a specific conditional independence constraint and learn the network parameters by setting an unbiased estimate of the true mean-squared error (MSE) of the mappings as an optimizing objective.…”

A Non-Linear Convolution Network for Image Processing

“…The literature provides several approaches focused on neural network architectures that exploit non-linear elements different from the canonical activation functions. Some of these methods express the non-linearity in a punctual way, for example by introducing a punctual multiplication operation on the outputs as in [2]. In [3], instead, the authors introduce Volterra kernels in convolutional neural networks, and in [4] a quadratic kernel is introduced too.…”

Object Detection Based on Global-Local Saliency Constraint in Aerial Images

“…Therefore, this article mainly discusses single-image processing. The development of deep neural networks has a significant boost for image denoising algorithms [56], [57], [58], [59], [60], [61], [62]. Most image denoising algorithms regard the denoising process as a process to minimize the empirical risk, given the corrupted inputsx i and the clean targets y i , then the process of training this regression model can be formulated as…”