Dual Frequency Transformer for Efficient SDR-to-HDR Translation

作     者：Gang Xu Qibin Hou Ming-Ming Cheng Gang Xu;Qibin Hou;Ming-Ming Cheng

作者机构：Tianjin Media Computing CenterCollege of Computer ScienceNankai UniversityTianjin300000China 

出 版 物：《Machine Intelligence Research》 (机器智能研究（英文版）)

年 卷 期：2024年第21卷第3期

页      面：538-548页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 080203[工学-机械设计及理论] 0835[工学-软件工程] 0802[工学-机械工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by National Natural Science Foundation of China(Nos.61922046 and 62276145) the National Key Research and Development Program of China(No.2018AAA0100400) Fundamental Research Funds for Central Universities,China(No.63223049) 

主　　题：Standard-dynamic-range to high-dynamic-range(SDR-to-HDR)translation Transformer dual frequency attention(DFA) frequency-aware feature decomposition efficient model 

摘      要：The SDR-to-HDR translation technique can convert the abundant standard-dynamic-range (SDR) media resources to high-dynamic-range (HDR) ones, which can represent high-contrast scenes, providing more realistic visual experiences. While recent vision Transformers have achieved promising performance in many low-level vision tasks, there are few works attempting to leverage Transformers for SDR-to-HDR translation. In this paper, we are among the first to investigate the performance of Transformers for SDR-to-HDR translation. We find that directly using the self-attention mechanism may involve artifacts in the results due to the inappropriate way to model long-range dependencies between the low-frequency and high-frequency components. Taking this into account, we advance the self-attention mechanism and present a dual frequency attention (DFA), which leverages the self-attention mechanism to separately encode the low-frequency structural information and high-frequency detail information. Based on the proposed DFA, we further design a multi-scale feature fusion network, named dual frequency Transformer (DFT), for efficient SDR-to-HDR translation. Extensive experiments on the HDRTV1K dataset demonstrate that our DFT can achieve better quantitative and qualitative performance than the recent state-of-the-art methods. The code of our DFT is made publicly available at https://***/CS-GangXu/DFT.