基于密集连接编解码网络的HDR重建算法

HDR reconstruction algorithm based on densely connected encoder decoder networks

针对动态场景下高动态范围(High Dynamic Range,HDR)重建会出现伪影现象的问题,提出一种基于密集连接编解码网络的HDR重建算法。该算法基于通道和空间维度的注意力机制模块,根据注意力特征图引导并提取不同曝光低动态范围图像相对应的图像特征,自动抑制由失准和过饱和引起的不期望分量,增强非参考图像中所期望的精细细节。通过密集连接编解码网络模块,在编码部分逐步提取特征图的多层次细节,并利用密集连接增强信息交互,使用解码器还原图像细节特征,最终生成少伪影、高质量的高动态范围图像。实验结果表明,该算法可以有效减弱伪影现象,得到细节丰富的高动态范围图像。

For the problem that the high dynamic range(HDR)reconstruction in dynamic scenarios will have artifacts,an HDR reconstruction algorithm based on a densely connected coder-decoder network is proposed.The algorithm is based on the attention mechanism module in channel and spatial dimensions,which guides and extracts image features corresponding to low dynamic range images with different exposures according to the attention feature map,automatically suppresses undesired components caused by misalignment and oversaturation,and enhances the desired fine details in the non-reference image.The multi-level details of the feature map are gradually extracted in the coding part through the densely connected coder-decoder network module,and dense connection is adopted to enhance the information interaction,and the image detail features are restored using the decoder,which ultimately generates artifact-less and high-quality high-dynamic-range images.Experiment results show that the algorithm can effectively attenuate the artifact phenomenon and obtain high dynamic range images with rich details.