Loading [MathJax]/extensions/MathZoom.js
Consumer Camera Demosaicking and Denoising With a Collaborative Attention Fusion Network | IEEE Journals & Magazine | IEEE Xplore

Consumer Camera Demosaicking and Denoising With a Collaborative Attention Fusion Network


Abstract:

For the consumer cameras with Bayer filter array, raw color filter array (CFA) data collected in real-world is sampled with signal-dependent noise. Various joint denoisin...Show More

Abstract:

For the consumer cameras with Bayer filter array, raw color filter array (CFA) data collected in real-world is sampled with signal-dependent noise. Various joint denoising and demosaicking (JDD) methods are utilized to reconstruct full-color and noise-free images. However, some artifacts (e.g., remaining noise, color distortion, and fuzzy details) still exist in the reconstructed images by most JDD models, mainly due to the highly related challenges of low sampling rate and signal-dependent noise. In this paper, a collaborative attention fusion network (CAF-Net), with two key modules, is proposed to solve this issue. Firstly, a multi-weight attention module is proposed to efficiently extract image features by realizing the interaction of spatial, channel, and pixel attention mechanisms. By designing a local feedforward network and mask convolution aggregation of multiple receptive fields, we then propose an effective dual-branch feature fusion module, which enhances image details and spatial correlation. Accordingly, the proposed two modules significantly facilitate our CAF-Net to recover a high-quality image, by accurately inferring the correlations of color, noise, and the spatial distribution of the CFA data. Extensive experiments on demosaicking, synthetic, and real image JDD tasks prove that the proposed CAF-Net can achieve advanced performance in terms of objective evaluation index metrics and visual perception.
Published in: IEEE Transactions on Consumer Electronics ( Volume: 70, Issue: 1, February 2024)
Page(s): 509 - 521
Date of Publication: 12 December 2023

ISSN Information:

Funding Agency:


I. Introduction

Modern consumer-grade camera sensors capture only one single color at each photoreceptor. The camera color filter array (CFA) defines this capture pattern and results in a regular subsampling of red (R), green (G), and blue (B) color channels. The process of recovering missing color components from mosaic CFA data is called color demosaicking. Due to hardware limitations, noise may be introduced at different stages when converting photos to digital pixel values. The noise depends on various factors such as scene and sensor, which makes the noise signal-dependent [1]. Thus, each pixel of the captured raw image contains only partial information necessary to reconstruct the full sample, and the noise further corrupts the captured data.

References

References is not available for this document.