# Image Scaling using Deep Convolutional Neural Networks

DevFeed: [Image Scaling using Deep Convolutional Neural Networks](<https://devfeed.tech/articles/image-scaling-using-deep-convolutional-neural-networks-31890.md>)

Original publisher: [Read original article](<http://engineering.flipboard.com//2015/05/scaling-convnets>)

Author: https://twitter.com/normantasfi (Norman Tasfi)

Published: 2015-05-06T00:00:00Z

Content type: article

Language: en

Sources: [Flipboard](<https://devfeed.tech/sources/flipboard.md>)

Topics: [Deep neural networks](<https://devfeed.tech/topics/deep-neural-networks.md>), [Machine Learning & Artificial Intelligence](<https://devfeed.tech/topics/machine-learning-artificial-intelligence.md>), [Image](<https://devfeed.tech/topics/image.md>), [Algorithms](<https://devfeed.tech/topics/algorithms.md>), [Web](<https://devfeed.tech/topics/web.md>), [Mobile](<https://devfeed.tech/topics/mobile.md>)

Tags: [algorithms](<https://devfeed.tech/tags/algorithms.md>), [architectures](<https://devfeed.tech/tags/architectures.md>), [deep](<https://devfeed.tech/tags/deep.md>), [image](<https://devfeed.tech/tags/image.md>), [machine-learning](<https://devfeed.tech/tags/machine-learning.md>), [mobile](<https://devfeed.tech/tags/mobile.md>), [neural-networks](<https://devfeed.tech/tags/neural-networks.md>), [pixel](<https://devfeed.tech/tags/pixel.md>), [presentation](<https://devfeed.tech/tags/presentation.md>), [web](<https://devfeed.tech/tags/web.md>)

## AI overview

This engineering article examines image upscaling for Flipboard and introduces convolutional neural networks alongside traditional interpolation methods. It explains how enlarging low-resolution images can produce pixelation, smoothing, noise, haloing, and other artifacts, and outlines the model discussion, preliminary results, design decisions, and possible product applications.

## Source excerpt

This past summer I interned at Flipboard in Palo Alto, California. I worked on machine learning based problems, one of which was Image Upscaling. This post will show some preliminary results, discuss our model and its possible applications to Flipboard's products. High quality and a print-like finish play a key role in Flipboard's design language. We want users to enjoy a consistent and beautiful experience throughout all of Flipboard's content, as if they had a custom print magazine in hand. Providing this experience consistently is difficult. Different factors, such as image quality, deeply affect the overall quality of the presented content. Image quality varies greatly depending on the image's source. This varying image quality is especially apparent in magazines that display images across the whole page in a full bleed format. When we display images on either the web or mobile devices they must be above a certain threshold to display well. If we receive a large image on our web product we can create breathtaking full bleed sections. Full bleed High Quality Image Lower resolution images introduce pixelation, over smoothing and artifacts when scaled above 100%. This is especially apparent in a full bleed presentation as seen below. This severely reduces the quality of presentation in our products. Full bleed Low Quality Image What is the cause of this? In general, when we need an image of size X that is required to be of size Y it must be run through a scaling algorithm. This algorithm performs a mathematical operation to scale the image pixels to the desired size Y. Some of the possible algorithms are bicubic, bilinear, and nearest-neighbor interpolation. Many of the algorithms listed above perform an interpolation between pixel values to create a transition. These algorithms use the surrounding pixels to guess what the missing values should be in the new image. The problem in the case of scaling the image to a larger size is when there are too many 'new' values