Burst Image Super-Resolution with Base Frame Selection

CVPR 2024 Workshop (9^th NTIRE)

Sanghyun Kim, Min Jung Lee, Woohyeok Kim, Deunsol Jung, Jaesung Rim, Sunghyun Cho, Minsu Cho

Pohang University of Science and Technology (POSTECH)

Abstract

Burst image super-resolution has been a topic of active research in recent years due to its ability to obtain a high resolution image using complementary information between multiple frames in the burst. In this work, we explore using burst shots with non-uniform exposures to confront real-world practical scenarios by introducing a new benchmark dataset, dubbed Non-uniformly Exposed Burst Image (NEBI), that includes the burst frames at varying exposure times to obtain a broader range of irradiance and motion characteristics within a scene. As burst shots with non-uniform exposures exhibit varying levels of degradation, fusing information of the burst shots into the first frame as a base frame may not result in optimal image quality. To address this limitation, we propose a Frame Selection Network~(FSN) for non-uniform scenarios. This network seamlessly integrates into existing super-resolution methods in a plug-and-play manner with low computational cost. The comparative analysis reveals the effectiveness of the non-uniform setting for the practical scenario and our FSN on synthetic-/real- NEBI datasets.

Motivation

Why do we need non-uniformly exposed burst?

When burst photography uses the same exposure times for all frames, it often results in poor quality images due to camera noise and motion blur if the exposure time isn't optimal. It's hard to determine the best exposure time in real-world scenarios, making this approach less practical. However, using burst shots with varied exposure times can help reconstruct high-resolution images as though the optimal exposure time was used.

Why do we need a Frame Selection Network?

Burst shots with non-uniform exposures display varying degrees of quality, affecting the alignment and fusion of features between the base and subsequent frames, which can reduce overall image quality. Previous methods typically use the first frame as the base frame, ignoring its potential negative effects on image restoration. To address this issue, we propose a Frame Selection Network (FSN) that identifies the most suitable base frame to enhance overall image quality.

Method

FSN first constructs the image feature of each frame using CNN. The constructed image feature is fed into our Correlation-based Motion Aware (CMA) blocks, which update the image feature based on motion information extracted from the burst frames. To extract motion information from the burst frames, we propose Feature Correlation Module (FCM), which computes the local correlation along both spatial and temporal axes.

Experiments

Evaluation on Synthetic-/Real-NEBI

Comparison of existing burst super-resolution models and their variants incorporating our Feature Selection Network (FSN), along with an evaluation against the Auto-Exposure (AE) algorithm.

Training strategies on Real-NEBI

When training a Frame Selection Network (FSN) with different targets like PSNR, SSIM, and LPIPS using synthetic data, and then evaluating it on real data, the training target affects the performance.

Ablation study

Comparison of the performance of a baseline super-resolution network against variants that incorporate varying numbers of CMA blocks.

Qualitative results

BIPNet, when combined with our frame selector, improves the clarity of high-frequency image details by effectively merging complementary information from multiple frames into the chosen base frame. Conversely, without the frame selector, BIPNet tends to produce blurry images due to significant degradation in the initial frame.