TOTTORI UNIVERSITYPattern Recognition System Laboratory
Menu

RESEARCH

Facial Deblur Inference using Subspace Analysis for Recognition of Blurred Faces

Key Points

We use statistical models learned from blurred face images to estimate the point spread function that describes the blur in an input image. We then use the estimated function to remove the blur and obtain an image suitable for face recognition.

blur-method.png

Abstract

A face blurred by poor focus or movement can be difficult to recognize. This study estimates how a face image was blurred, removes the blur, and then performs face recognition. The pattern describing how image details spread out is called a point spread function. Estimating it correctly from just one blurred image is difficult.

We therefore learn from face images whose blur is already known. We organize their features so that faces with the same type of blur are similar, and build a model for each predefined set of blur patterns. For an unfamiliar image, we find the closest model and use its corresponding blur pattern to restore the image.

In experiments with artificially added focus and motion blur on the FERET face database, recognition improved substantially over the existing comparison methods. We also observed improvement on genuinely blurred images in the FRGC 1.0 database. Performance improved further when we combined the method with local phase quantization, or LPQ, another technique that extracts features from small regions of an image.

Publications

Publication (Japanese)

otama1.png
Back to research