High-dynamic Range Video Generation Method Based On Adaptive Priors Dynamic Mapping

Contemporary displays with the capability to present ultra-high-definition media content, featuring high dynamic range(HDR) and wide color gamut, are increasingly accessible, yet their overall availability remains limited. Considering the vast volume of standard dynamic range(SDR) media content that currently exists, there is a pressing need for suitable conversion techniques. This paper introduces a novel network called the adaptive priors dynamic mapping network(APDMNet), which enables end-to-end adaptive transformation of SDR frames into HDR frames, drawing inspiration from the global attention mechanism and dynamic convolution kernels. The proposed approach comprises a basic network and an adaptive priors network. The basic network encompasses global mapping, local mapping, and detail refining, and is responsible for feature mapping and image processing. The adaptive priors network extracts global prior and attention prior information from the input, integrating it into the basic network to facilitate adaptive feature mapping. Moreover, to address the inconsistency between the clipping training and the global test in the image recovery task, we introduce a local transformation mechanism during testing. This allows the global pooling operation in the adaptive priors network to produce improved results. Experimental results on the HDRTV1K dataset demonstrate that APDMNet outperforms previous methods in terms of performance, significantly improving the perceived quality of the generated HDR frame.