PROJECT TITLE :

Adaptive Residual Networks for High-Quality Image Restoration - 2018

ABSTRACT:

Image restoration ways primarily based on convolutional neural networks have shown nice success in the literature. However, since most of networks are not deep enough, there's still some space for the performance improvement. On the other hand, though some models are deep and introduce shortcuts for straightforward coaching, they ignore the importance of location and scaling of different inputs within the shortcuts. Thence, existing networks will solely handle one specific image restoration application. To handle such issues, we tend to propose a completely unique adaptive residual network (ARN) for high-quality image restoration in this Project. Our ARN may be a deep residual network, which is composed of convolutional layers, parametric rectified linear unit layers, and some adaptive shortcuts. We tend to assign totally different scaling parameters to completely different inputs of the shortcuts, where the scaling is taken into account as part parameters of the ARN and trained adaptively according to totally different applications. Due to the special construction of ARN, it will solve many image restoration issues and have superior performance. We have a tendency to demonstrate its capabilities with 3 representative applications, together with Gaussian image denoising, single image super resolution, and JPEG image deblocking. Experimental results prove that our model greatly outperforms various state-of-the-art restoration ways in terms of both peak signal-to-noise ratio and structure similarity index metrics, e.g., it achieves 0.2-zero.3 dB gain in average compared with the second best technique at a wide range of situations.