A Deep Learning based magnifying glass

At idealo.de we trained a state-of-the-art Deep Convolutional Neural Network to 🔎super-scale🔎 small images.

Francesco Cardinale
Dec 5, 2018 · 9 min read
Super scaling a low-resolution butterfly.

Overview

Training

ISR training flow.

Loss

Evaluation

Setup

Tensorboard graphs of the 90 training epochs. Top-left: training loss that is back-propagated to the network at the end of each epoch. Top-right: the loss on the out-of-training set that we used to keep track of generalization performance. Bottom left and right: the PSNR values for training and validation set, respectively.

Results

LR image (left), reconstructed SR (center), GIMP baseline scaling (right). Source: DIV2K dataset.

Understanding the results

Heatmap for pixel-wise HR-SR error. Darker colors mean higher error and lighter lower error, or better results.

A few words on non-standard ground truth in deep learning tasks

Pictorial description of the different feeding methods we tried.

Going further

Low resolution image of a sandal.
Super-scaled image of a sandal.

Links

idealo Tech Blog

👨‍💻👩‍💻 idealo tech and product folks writing about…