USIM-DAL: Uncertainty-aware Statistical Image Modeling-based Dense Active Learning for Super-resolution

TL;DR

USIM-DAL introduces an uncertainty-aware active learning framework for dense image regression tasks like super-resolution, reducing annotation costs and improving accuracy by leveraging probabilistic deep neural networks and image statistics.

Contribution

The paper presents USIM-DAL, a novel active learning framework that models uncertainty in dense regression tasks using probabilistic neural networks and image statistics, addressing a key research gap.

Findings

USIM-DAL outperforms existing dense regression active learning methods.

The framework effectively reduces annotation costs while maintaining high accuracy.

Demonstrated success across diverse datasets including natural, medical, and remote sensing images.

Abstract

Dense regression is a widely used approach in computer vision for tasks such as image super-resolution, enhancement, depth estimation, etc. However, the high cost of annotation and labeling makes it challenging to achieve accurate results. We propose incorporating active learning into dense regression models to address this problem. Active learning allows models to select the most informative samples for labeling, reducing the overall annotation cost while improving performance. Despite its potential, active learning has not been widely explored in high-dimensional computer vision regression tasks like super-resolution. We address this research gap and propose a new framework called USIM-DAL that leverages the statistical properties of colour images to learn informative priors using probabilistic deep neural networks that model the heteroscedastic predictive distribution allowing uncertainty quantification. Moreover, the aleatoric uncertainty from the network serves as a proxy for error that is used for active learning. Our experiments on a wide variety of datasets spanning applications in natural images (visual genome, BSD100), medical imaging (histopathology slides), and remote sensing (satellite images) demonstrate the efficacy of the newly proposed USIM-DAL and superiority over several dense regression active learning methods.