•  
  •  
 

Document Type

Article

Keywords

Grey video, ELU, reference image, Trilinear up sampling, encoder and decoder, BN, Lab, RGB, PSNR, SSIM

Abstract

Colorizing grayscale videos is a challenging task that involves adding colors to monochromatic videos to make them appear as natural and realistic as possible, despite the absence of information about the original color distribution. This task is crucial for applications such as restoring historical videos and creating media content requiring realistic and accurate colorization. However, existing methods often face issues such as poor temporal stability, color distortions, and the need for extensive post processing under certain conditions. To address these challenges, this study proposes a novel approach comprising a pre-processing network and a source-reference network trained in an end-to-end manner. The pre-processing network employs an encoder-decoder architecture enhanced with temporal convolutions and skip connections, enabling it to improve video quality, adapt to resolution changes, and leverage batch normalization (BN) and exponential linear unit (ELU). The source-reference network incorporates an encoder for reference image processing and a fusion module with residual blocks to combine feature maps through a source-reference attention mechanism. The final output is generated in the Lab color space and converted to the RGB format, ensuring high-quality video colorization with enhanced temporal stability. Experimental evaluations demonstrate that the proposed model achieves significant improvements in color realism and temporal stability, with a PSNR of 37.89 and an SSIM of 0.999982—which surpasses those of most state-of-the-art methods. These results confirm the effectiveness and applicability of the proposed method for video colorization tasks, making it a robust solution in the domain.

Download

Share

COinS