Q-TrHDRI: A Qurey-Based Transformer for High Dynamic Range Imaging with Dynamic Scenes

Abstract

In the absence of well-exposed contents in images, high dynamic range image (HDRI) provides an attractive option that fuses stacked low dynamic range (LDR) images into an HDR image. Existing HDRI methods utilized convolutional neural networks (CNNs) to model local correlations, which can perform well on LDR images with static scenes, but always failed on dynamic scenes where large motions exist. Here we focus on the dynamic scenarios in HDRI, and propose a Query-based Transformer framework, called Q-TrHDRI. To avoid ghosting artifacts induced by moving content fusion, Q-TrHDRI uses Transformer instead of CNNs for feature enhancement and fusion, allowing global interactions across different LDR images. To further improve performance, we investigate comprehensively different strategies of transformers and propose a query-attention scheme for finding related contents across LDR images and a linear fusion scheme for skillfully borrowing complementary contents from LDR images. All these efforts make Q-TrHDRI a simple yet solid transformer-based HDRI baseline. The thorough experiments also validate the effectiveness of the proposed Q-TrHDRI, where it achieves superior performances over state-of-the-art methods on various challenging datasets.