无人机夜间场景低照度图像增强方法

Low-light image enhancement method for UAV in night scenes

针对无人机夜间场景低照度图像增强的问题,提出一种简单有效的解决方案,无需任何与任务相关的数据。该方法遵循图像自回归原理和灰度世界色彩恒定假说,通过构造RGB通道的高斯分布N(ηi,σi)采样噪声,构建由5层卷积网络组成的超轻量自回归模型,实现低照度图像的高质量增强。实验结果表明:所提方法在低照度图像增强方面具有较强的竞争力,能够增强图像的亮度和细节信息,获得良好的视觉效果。最重要的是,该模型非常轻量化,毫秒级的推理速度适合部署到无人机实现夜间场景低照度图像的高质量增强。同时该方法基于零样本学习,无需训练数据,具有良好的泛化性。

A simple and effective solution without task-relevant data is proposed,aiming at the problem of low-light image enhancement for unmanned aerial vehicle(UAV)in night scenes.The method follows the image autoregressive principle and the grey-world color constancy hypothesis.It achieves high-quality enhancement of low-light images by constructing a Gaussian-distributed N(ηi,σi)sampling noise of the RGB channel and training an ultra-lightweight autoregressive model consisting of a five-layer convolutional network.The experimental results indicate that the proposed method is highly competitive in low-light image enhancement,as it enhances brightness and detail information and achieves good visual effects.Notably,the model is lightweight,and the millisecond-level inference speed is suitable for high-quality image enhancement of UAVs in low-light night scenes.Moreover,the proposed method is based on zero-sample learning,which requires no training data,thereby has good generalisation.