空间光学技术研究室知识库

Low-light remote sensing is a powerful complement to daytime optical remote sensing and can greatly expand the time domain of high-resolution earth observations, and make day and night imaging possible. However, when a low-light sensor is used in the morning dusk and dawn, the captured images have characteristics of low contrast, low brightness, and low signal-to-noise ratio, which severely restrict the identification and interpretation of ground objects. Traditional low-light image enhancement algorithms such as histogram equalization, gamma conversion, and contrast-limited adaptive histogram equalization algorithm, and so on can enhance the low-light remote sensing image and solve the problem of contrast enhancement, but the noise amplification effect brought by the enhancement will degrade the signal-to-noise ratio of the enhanced image. Therefore, in this paper, a data-driven low-light remote sensing image enhancement algorithm is studied. First of all, lots of low-light raw image data pairs corresponding to very low illumination are captured. Then, these raw image data are used to train a deep fully convolutional neural network composed of an encoder–decoder structure. After that, the low-light remote sensing images could be enhanced by the pretrained net structure. The numerical results demonstrate that the fully convolutional neural network based on enhancement algorithm greatly improves the brightness and the contrast of low-light images compared with the traditional enhancement algorithms while a high enough signal-to-noise ratio could be preserved, which will make interpretation and identification much easier. © 2019, Springer Nature Singapore Pte Ltd.