Unsupervised conversion method of high bit-depth remote sensing images using contrastive learning

Abstract

Currently, remote sensing images are frequently stored in high bit-depth formats exceeding 10 bits. However, the standard 8-bit format remains the fundamental data format for visualization and deep learning applications. Traditional methods typically rely on manually adjusting the parameter threshold of the tone mapping operator to obtain 8-bit images, resulting in low automation. Although tone mapping methods based on deep learning have gradually supplanted traditional techniques, but such methods are mainly aimed at natural scene images taken by digital cameras. There are problems such as incompatibility between data format and image semantics, and it is difficult to meet the scale dependence of remote sensing image applications. To address these challenges, we propose an unsupervised bit-depth conversion method for remote sensing images that integrates generative adversarial networks with contrastive learning. We draw an analogy between gray value mapping and the motion of thermal field particles, constructing a transformer generator based on thermodynamic principles. Leveraging the analogous characteristics of high and low bit-depth image histograms, we introduce a histogram shape context contrastive loss to regulate the color distribution of the generated images. Furthermore, in light of the large-scale application characteristics of remote sensing images, we propose a post-processing method based on hybrid histogram matching to enhance image quality while generating seamless whole-scene images. We developed relevant datasets and conducted experiments, with results demonstrating that the proposed method achieves superior bit-depth conversion effects compared to existing methods. Code and data can be found at https://github.com/ZzzTD/Bit-depth_conversion.