The impact of Single Image Super Resolution techniques

The impact of Single Image Super. Resolution techniques on Face Recognition tasks Antônio Luís Sombra de Medeiros 07/2019

Research Question • Can Face Recognition tasks be improved using Super-Resolution tehcniques? • What are the impacts of different Super. Resolution techniques on the performance of face recognition tasks?

Face Recognition • Face recognition is of great importance in many computer vision applications, such as human-computer interactions, Security systems, Military and Homeland Security. • • • Payment Access security Criminal identification Advertising Healthcare

Problem • face recognition systems mostly work with imagesvideos of proper quality and resolution. • In videos recorded by surveillance camera, due to the distance between people and cameras, people are pictured very small and hence challenge face recognition algorithms. • Database samples of HR and test images of LR

Problem

Super-Resolution • Improves human interpretation • The application of SR techniques covers a wide range of purposes such as Surveillance video, Remote sensing, Medical imaging (CT, MRI, Ultrasound).

Traditional x Deep Learning based SR models

Very Deep Super-Resolution netwok (VDSR)

Super-Resolution GAN (SRGAN)

SRGAN loss function

Super-Resolution training - LR data shyntesis

Degradation Model

Dataset

Test Results

Ground truth Bicubic upsampled Results VDSR SRGAN

Face Recognition

Model • Trained Resnet based model from dlib. • “Res. Net network with 29 conv layers. It's essentially a version of the Res. Net-34 network from the paper Deep Residual Learning for Image Recognition by He, Zhang, Ren, and Sun with a few layers removed and the number of filters per layer reduced by half” • Trained on approx. 3 million images of faces from internet.

Dataset for Face Recognition Evaluation

Face recognition results for LR images

Face recognition results using SR images

Conclusion • We tested two new SR techniques on face recognition application for LR resolution images • Relevant improvement using SR • SRGAN obtained best results

THANKS!