Deep Learning

The evolution of deep learning has been led by applications in computer vision, image processing, computer graphics, and computational photography such as object detection and recognition, image restoration, and image synthesis and manipulation. In our lab, we study deep learning such as new neural network architectures, and training strategies for improving the performance of such applications.

Blur map estimation. Left: input images with defocus and motion blur. Right: their corresponding blur map (red: defocus blur, blue: motion blur, black: no-blur) (Kim et al., Pacific Graphics 2018)

Neural network architecture for blur map estimation (Kim et al., Pacific Graphics 2018)


Single image super-resolution (left: low-resolution image upsampled by bicubic interpolation, right: single-image super-resolution result using a GAN based CNN) (Park et al., ECCV 2018)

Neural network arcthiecture for single image super-resolution (Park et al., ECCV 2018)


Image Deblurring

Blur, which is one of the most annoying artifacts in photographs, is caused by physical limitations of cameras such as long exposure time or lens aberration. Deblurring is a problem to remove blur from a blurry image so that valuable hidden information can be revealed. Deblurring can benefit consumer cameras, medical imaging, aerial and satellite imaging, and many others.