基于多残差注意力深度收缩网络的超微光图像增强方法

An extreme low-light imaging enhancement method based on a multi-residual attention shrinkage network

超微光成像可在极度黑暗的环境中给观察者提供近乎白昼的视觉体验,在许多民用和军事应用中起着至关重要的作用。超微光环境下拍摄的图像和视频通常存在亮度与对比度极低、噪声水平高、场景细节和色彩严重缺失等固有缺陷,近年来,深度学习为超微光成像的研究带来了新的机遇。文中采集并提供了一组实用性更强的超微光训练数据集,提出了一种多残差注意力深度收缩网络(Multi Residual Attention Shrinkage Network),以此实现了一种新的超微光成像方法。通过成功研制的小型化样机证实了该方法的工业量产前景。实现了基于通道注意力和空间注意力的残差内注意力机制,以及基于深度软阈值收缩的外注意力机制,不仅可以有效提取并还原极低照度环境下的图像细节信息,恢复场景真实色彩,而且可以有效去除此类环境下由成像设备感光不足带来的巨量噪声。实测效果显示该方法可对极低照度环境进行有效的增强且实时性高。通过与多种业界最新方法比较,文中方法在主观视觉体验以及客观参数两方面均表现更好。

Extreme-low light imaging enhancement can provide observers with near-daylight visual experience in extremely dark environments,and plays a vital role in many civil and military applications.Images and videos taken in ultra-low light level environment usually have inherent defects,such as extremely low brightness and contrast,high noise level,and serious lack of scene details and colors.In recent years,deep learning has brought new opportunities for the research of ultra-low light level imaging.In this paper,we collect and provide a series of more practical ultra-low light level training data sets,and propose a multiple residual attention shrinkage network.Thus,a new ultra-low light level imaging method is developed.The prospect of industrial mass production of this method is confirmed by the successfully developed miniaturized prototype.This paper implements the residual internal attention mechanism based on channel attention and spatial attention,as well as the external attention mechanism based on depth soft threshold shrinkage.This approach can not only effectively extract and restore the image detail information under extremely low illumination environment,restore the true color of the scene,but also effectively remove the huge amount of noise caused by insufficient light sensitivity of imaging equipment in such environment.The measured results show that this method can effectively enhance the extremely low illumination environment and has high real-time performance.Compared with the latest methods in the industry,the proposed method is superior in terms of subjective visual experience and objective parameters.