Highly compressed images from surveillancecamera systems suffer from block noise. We propose two methods to restore such degraded face images based on an example-based Super-Resolution (SR) method. The base method preliminarily generates a database of patches from a training set of face-image examples and reconstructs a High-Resolution (HR) image from a given Low-Resolution (LR) image by using the patches and taking the positions of facial parts into consideration. The proposed methods aim to

more »

... store highly-compressed degraded images instead of LR images. One of the proposed methods named as the Direct method, synthesizes a restored texture directly with a given degraded image. The other method, named as the Smooth method, synthesizes a restored texture with a filtered image generated from a given degraded image. The Direct method results in a 3.2 dB improvement in terms of Peak Signal-to-Noise Ratio (PSNR) on average for the lowest quality rate i.e. 1% (around 120:1 compression rate), while the conventional Gaussian-filtering method results in a 2.5 dB improvement. Although the Direct method results in better quality for highly-compressed images compared to the conventional Gaussian-filtering method and Smooth method, unnatural block noise is still observed in the restored images. The Smooth method yields more natural images and better PSNRs for images when the quality rate is around 5%.  Index Terms-example-based, compressed image, image restoration, Gaussian-filtering, Direct method, Smooth method 9