Face Image Super-Resolution using Deep Learning

By DeepStrive

Introduction


High-resolution human face images are important in various fields. Particularly, in security and surveillance fields, high-resolution face images are required for face recognition or detection purpose. The multimedia industry also desires high-resolution face images since they provide better audience perception and satisfaction.

In reality, resolution of face images may not necessarily be as high as desired. These may sometimes be limited by hardware of imaging devices , or the fact that the image is sometimes out of focus. Image resolution may also be limited by the distance between the human subject and the imaging device, such that the face image taken is small. Therefore, it is highly desirable and valuable if high-resolution face image could be reconstructed from low-resolution image using super-resolution techniques. The latest advancement in deep learning technology opens possibility for more plausible image super-resolution by learning from realistic facial features.


Objectives


This project aims at implementing a deep neural network for super resolution with a narrower input scope, which is face images. By experimenting various deep learning algorithms and improving existing algorithms, we expect that our model will produce photo-realistic outputs while reducing computation power and being able to handle multi-scale super-resolution. Also, the output image is expected to be a high resolution face image of the same person.

There are several rooms for improvements on existing solutions of super-resolution by deep learning and algorithmic models and algorithmic models. In terms of:

  1. Accuracy and perceptual quality of outputs
  2. Computational resources required, such as time and memory usage
Moreover, existing solutions for super-resolution are usually more generic, without specializing on a category of images. We believe that studying face image super-resolution could bring new value to the field of image super-resolution.

Schedule


Task Estimated start time Estimated completion time Completion status
Stage 1
Initial self-study on deep learning Aug 20, 2018 Oct 15, 2018 80%
Image collection and pre-processing Sep 1, 2018 Sep 30, 2018 100%
Literature Review Sep 1, 2018 Sep 30, 2018 100%
Phase 1 deliverable: Detailed project plan Sep 1, 2018 Sep 30, 2018 100%
Phase 1 deliverable: Project webpage Sep 1, 2018 Sep 30, 2018 100%
Stage 2
Get familiar with tools and environment Sep 15, 2018 Oct 15, 2018 50%
Replication of results from research papers Oct 1, 2018 Oct 31, 2018 0%
Stage 3
Model creation, evaluation and implementation
sDeliverable: Deep learning model – initial version
Nov 1, 2018 Jan 6, 2018 0%
First presentation Jan 7, 2019 Jan 11, 2019 0%
Phase 2 deliverable: Interim Report Dec 24, 2018 Jan 20, 2019 0%
Enhance deep learning model
Deliverable: deep learning model – enhanced version
Jan 21, 2019 Feb 28, 2019 0%
Stage 4
Implement optional features
Deliverable: deep learning model – extended version
Mar 1, 2019 Apr 7, 2019 0%
Phase 3 deliverable: Implementation and Final Report Mar 18, 2019 Apr 14, 2019 0%
Final presentation Apr 15, 2019 Apr 19, 2019 0%
Project exhibition Apr 15, 2019 Apr 29, 2019 0%
Project competition (if selected) Apr 30, 2019 May 29, 2019 0%

CONTACT

Dr. Kenneth Wong

Dr. Kenneth Wong

Supervisor
Florence Tsang

Florence Tsang

Year 4, Computer Science
Frankie Lo

Frankie Lo

Year 4, Computer Science
Willis Wong

Willis Wong

Year 4, Computing and Data Analytics