Deep bilateral learning for image interpolation

Jiahuan Ji

doi:10.1016/j.knosys.2026.116266

Deep bilateral learning for image interpolation

2026-07-19

0

PRE

AI

DOI：10.1016/j.knosys.2026.116266

Original

Original request for help

Save

Abstract

En 中文

Among the existing deep learning-based image interpolation methods, a common practice for estimating the value of each missing pixel is to estimate the intensity difference between it and its associated known or anchor pixel from the same interpolation window. The predicted difference is then added to the value of the anchor pixel as the final predicted pixel value. However, we have observed that this approach inevitably results in inaccurate estimation when the missing pixel is an 'outlier' with respect to the anchor pixel. To address this issue, a novel deep bilateral learning (DBL) method is proposed in this paper, which is motivated by the fundamental idea of the classical bilateral filter. For each missing pixel to be estimated, our DBL is designed to learn two attributes between it and the anchor pixels from a pre-defined support region, including their intensity differences and a set of distance-based spatial weights. The former is added to the respective anchor pixel values to yield a set of candidate pixel values for the missing pixel. The spatial weights are then used to compute a weighted sum of these candidates, yielding the final estimated value. To verify the effectiveness of our DBL, a DBL-based network, called the deep bilateral learning network (DBLN), is designed for image interpolation. Extensive experimental results have clearly shown that our DBLN can produce superior performance compared to a number of state-of-the-art image interpolation methods.

Keywords:

Image interpolation

Deep bilateral learning

Convolutional neural network

Deep bilateral learning for image interpolation

Abstract

Journal

Organization

Cited Papers