SRN-Deblur

Why we included this project

Camera shake and motion blur are the usual culprits behind ruined photos, and this repo holds the official implementation of a CVPR 2018 method that tackles them in a single image. The scale-recurrent network works coarse-to-fine, processing the image at several resolutions so that a blurry input gets progressively sharpened. You can run it right away: pretrained checkpoints are included, and run_model.py takes a folder of images and writes deblurred outputs without any retraining. The catch is that it's 2018 research code, so you'll need Python 2.7 and TensorFlow 1.4, and input dimensions must be multiples of 16; that makes it a better fit for teams comfortable adapting older code than for anyone expecting a drop-in modern library. If you're studying or reproducing classic deblurring architectures, the training pipeline against the GoPro dataset is documented here as well.