VQALS: A Video Question Answering Method in Low-Light Scenes Based on Illumination Correction and Feature Enhancement

Abstract

In low-light scenes, videos often exhibit low brightness, leading to less evident details in regional features. The current video question answering models have made significant progress in the fusion and reasoning of cross-modal information. However, they perform poorly in effectively extracting useful information and salient features in low-light scenes. To tackle this challenge, we propose a video question answering method in low-light scenes, in which two modules are developed: illumination correction module and feature enhancement module. The illumination correction module enhances visual quality by applying adaptive enhancement to the video with a variational threshold, thereby extracting more feature information. The feature enhancement module further enriches and strengthens important information in the features by introducing a dynamic learning strategy to enhance spatial features by two branches, providing reasonable evidence for inferring the correct answer. Finally, the enhanced visual features are fused with question features to infer and generate proper answers. We perform extensive experiments on public datasets. The experimental results manifest the advantages and effectiveness compared with state-of-the-art methods in terms of accuracy in video question answering task.