Image Restoration Using Deep Neural Networks

We have presented a study of the MemNet[1] model which performs three tasks: JPEG deblocking, image denoising, and super resolution image generation. This model tackles the image restoration problem by using a recursive block which has both long term and short term memory blocks. These blocks help in achieving persistent memory in the model. We have also performed stress testing on the model. We have tested the model on different datasets for generalizing the model for different image restoration tasks.

The goal of our project is to develop an image processing pipeline which will be used for image restoration tasks - Noise removal, Super resolution and JPEG deblocking. This pipeline can be used as a pre-processing block with existing state-of-the-art object detection models such as Detectron[7], Fast R-CNN[14], Mask R-CNN[9]. We also wish to demonstrate the robustness of the model, and how well the model generalizes on image restoration tasks such as DeRain[5] and DeSnow[4].

Images are captured from areas ranging from professional photography to astronomy, surveillance, remote sensing, biomedical imaging etc. Interference in camera, varying lighting conditions like low light, extreme low light, different exposure ranges and external factors can often cause blurring and corruption of captured images. Image reconstruction has been a widely explored and researched topic in the field of computer vision. This problem statement is critical because this forms the premise for any of the aforementioned applications which need to make use of these images. It is highly essential that we remove the image corruption and improve the resolution of these images so that they can be efficiently used in ongoing research, academia and industrial applications.