“Image Super-Resolution Using Deep Convolutional Networks,” submitted to arXiv on December 31, 2014 by Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang, introduced SRCNN, the first method to apply a deep convolutional network end-to-end to the problem of single-image super-resolution - taking one low-resolution photo and producing a sharper high-resolution version. Before this, the state of the art used sparse-coding pipelines built from several separately engineered stages.
SRCNN learns a single mapping directly from low-resolution to high-resolution images, jointly optimizing every layer rather than tuning components in isolation. One of the paper’s elegant insights is that the older sparse-coding approach can be reinterpreted as a particular, shallow convolutional network, which framed the new deep method as a natural generalization rather than a break from prior work. The architecture is deliberately lightweight - only a few convolutional layers - yet it reached state-of-the-art quality at practical speed, and the authors extended it to handle color channels together.
The paper landed just as deep learning was sweeping computer vision after AlexNet, and it established the template that essentially all later neural super-resolution work followed: replace handcrafted multi-step pipelines with a learned end-to-end network. Two years later, SRGAN would push the same problem toward photorealistic texture using adversarial training.
Why business readers should care: super-resolution is one of the most quietly ubiquitous AI features in consumer products, powering phone camera zoom, video upscaling, and old-photo restoration. SRCNN is where the modern, learned version of that capability began.