Image Super-resolution

# Low-level Computer Vison

We presente a simple and effective framework for image super resolution, and propose a novel CNN with a self-feature based loss network that is capable of reconstructing images of pleasing visual quality. We speculate that keeping consistency of features and network will remove the artefacts caused by features from pre-trained networks. The proposed use of self-feature loss approach provides new insights into designing CNNs capable of performing optimally in wider application areas of imaging. A new insight to designing novel loss estimation functions for deep learning architectures.

Image Demosaicking

# Low-level Computer Vison

A fast and flexible network, namely DemoNet, is proposed for image demosaicking. By taking a tunable CFA pattern as input, DemoNet can be generalized to any CFA pattern with a simple modification. DemoNet exhibits perceptually appealing results with the end-to-end solution, without requiring extra aid to overcome the challenging artifacts, demonstrating its potential for pratical image demosaicking. We perform in-depth analysis on the impact of different guidance prior information of CFA patterns. Our findings provide empirical knowledge on CNN architecture design for future image demosaicking research.

Hard case demo: NO salient artifacts: aliasing, zippering, false colour, blur… most all current methods prone to them.

Over-exposure Correction

# Low-level Computer Vison

Over-exposure(OE) artifact occurs when a limited dynamic range of digital imaging device records the bright scene of which luminance level is higher than saturation level of the sensor. Correcting over-exposure aims to recover the fine details from the input. Existing methods are proposed mainly from the perspective of manually image pixel manipulation, therefore often come at the cost of high computational and algorithmic complexity. Recently, some deep learning based approaches on high dynamic range (HDR) reconstruction can be treated as image correction because their output HDR image can be tone mapped into the low dynamic range (LDR) image. However, existing tone mapping methods always fail to preserve the local details from the HDR domain when OE happens. In this paper, we present OEC, a convolutional neural network (CNN) for over-exposure correction. To our knowledge, it is the first framework capable of inferring photo-realistic natural images for over-exposure in LDR domain. To achieve this, we construct a training dataset of over-exposed and ground truth image pairs for end-to-end learning. Experimental results demonstrate the advantage of our method over existing OE correction methods with a significant margin.

The proposed CNN has Over-exposure Compensation input. In the figure below the images with different over-exposure are on the left. The exposure-compensated images are in the middle.

Image Denoising

# Low-level Computer Vison

A flexible denoising network, namely BiDNet, is proposed for non-blind and blind image denoising.To our best knowledge, this is the first network is adequate for those both cases and performance the state-of-the-art. We highlight CNNs are able to learn a direct mapping function instead of modeling image priors explicitly. Our model exhibits perceptually appealing results on both noisy images corrupted by AWGN and real-world noisy images, demonstrating its potential for practical image denoising.