Grayscale mapping of infrared images based on end-to-end deep neural networks

Abstract

Abstract The grayscale mapping of infrared images is an important research direction in the field of infrared imaging. A fast and scene-adaptive grayscale mapping method is crucial for visualizing high dynamic range original infrared images in various standard dynamic range output devices, such as printers and standard monitors. At present, mainstream grayscale mapping methods can only handle high dynamic range images in limited scenes and require extensive parameter adjustments to generate high-quality mapping results. In this paper, we propose a fast, parameter-free, and scene-adaptive grayscale mapping method to address this issue, which can achieve high subjective quality mapping results. Our model not only adapts to various categories of scenes, but also resolves the issues of insufficient contrast and significant loss of details in the grayscale mapping of high dynamic range infrared images. We explored the different impacts of the loss functions and normalization layers in the model on the mapping effect, and ultimately adopted L1 loss, perceptual loss, and batch normalization to accomplish our task. To ensure the production of high-quality mapping results, we used the objective metric of high dynamic range image quality assessment, specifically the tone mapping image quality index, to identify target images for training our model. We evaluated our results from both quantitative and qualitative perspectives, showcasing the high-quality output images generated by our model in a wide range of real-world scenarios. This substantiates the superiority of our approach.