基于多曝光图像生成的低照度图像增强

Low-light Image Enhancement Based on Multi-exposure Images Generation

  • 摘要: 低照度图像会使很多计算机视觉算法的鲁棒性降低,严重影响机器人领域的许多视觉任务,如自动驾驶、图像识别以及目标追踪等。为获取具有更多细节信息以及更大动态范围的增强图像,提出了一种基于多曝光图像生成的低照度图像增强方法。该方法通过分析真实拍摄的多曝光图像,发现不同曝光时长的图像的像素值之间存在线性关系,使得正交分解的思想可以应用于多曝光图像生成。多曝光图像是根据物理成像机制生成的,与真实拍摄图像更为相近。在将原图分解得到光照不变量和光照分量后,通过设计自适应算法生成不同的光照分量,再与光照不变量合成可以得到多曝光图像。最后利用多曝光图像融合方法获取具有更大动态范围的增强图像。该融合结果与输入图像保持一致,最终的增强图像可有效保留原始图像的色彩,自然度高。在真实拍摄的低照度图像公开数据集上进行了实验并与现有先进算法进行对比,结果表明,本文方法得到的增强图像与参考图像之间的结构相似性提高了2.1%,特征相似性提高了4.6%,增强图像与参考图像更接近且自然度更高。

     

    Abstract: Low-light images will cause the robustness degradation in many computer vision algorithms, which seriously affects various vision tasks in the context of robotics, such as automatic driving, image recognition and target tracking. In order to obtain the enhanced image with more details and a larger dynamic range, a low-light image enhancement method based on multi-exposure images generation is proposed. By analyzing the real-captured multi-exposure images, it is found that there is a linear relationship between pixels of the images with different exposure time, so the idea of orthogonal decomposition can be applied to generating multi-exposure images. Because the multi-exposure images are generated according to the physical imaging mechanism, they are similar to the real-captured images. After the original image is decomposed into an illumination invariant component and an illumination component, an adaptive algorithm is designed to generate various illumination components, then the multi-exposure images are generated by combining the various illumination components with the invariant one. Finally, a multi-exposure image fusion strategy is applied to obtaining the enhanced image with a larger dynamic range. The fusion result is consistent with the input images, and the final enhanced image can effectively retrain the colour of the original image with high naturalness. The proposed method is compared with the existing advanced algorithms through experiments on the public dataset of real-captured low-light images, and the results show that the structural similarity between the enhanced image and the reference image is improved by 2.1% by this method, the feature similarity is improved by 4.6%, and the image enhanced by this method is closer to the reference image and more natural than the others.

     

/

返回文章
返回