摘要
为解决现有的低照度图像增强方法存在的色彩失真、细节损失以及暗区增强不足和亮区增强过度导致低照度图像增强效果不理想的问题,提出了一种从局部到全局的零参考低照度图像增强方法。采用局部照度增强对低照度图像进行像素级增强,改进了自适应光照映射估计函数,提升了照度调整能力,避免了生成大量的迭代参数,提高了模型的推理速度;采用基于Transformer结构的全局图像调整对局部增强后的图像进行全局调整,解决了亮区照度增强过度的曝光问题和暗区照度增强不足的问题,提升了图像的整体对比度;优化损失函数,对低照度图像特征和增强图像特征进行相似性约束,提升了目标检测精度。实验结果表明,LOL数据集上的客观指标峰值信噪比和结构相似性达到了20.18 dB和0.80,MIT-Adobe FiveK数据集上达到了23.31 dB和0.87,ExDark数据集上增强后图像的目标检测精度提高了7.6%,有效提升了低照度图像可视化质量和目标检测效果。
To address the problem of unsatisfactory low-light image enhancement caused by color distortion,loss of details,insufficient enhancement of dark areas,and excessive enhancement of bright areas in existing low-light image enhancement methods,a zero-reference low-light image enhancement method from local to global is proposed.Initially,the low-illumination image is enhanced at the pixel level through local illumination enhancement,the adaptive light mapping function is improved,and the illumination adjustment ability is enhanced to avoid the generation of massive iterative parameters and improve the inference speed of the model.Subsequently,the global adjustment is performed on the locally enhanced image through Transformer-based global image adjustment to tackle the exposure problem of excessive illumination enhancement in the bright area and insufficient illumination enhancement in the dark area,thereby enhancing the overall contrast of the image.The loss function is optimized to constrain similarities between features of the low-illumination image and those of the enhanced image,thereby improving the target detection accuracy.The experimental results show that the objective metrics peak signal-to-noise ratio and structural similarity can reach 20.18 dB and 0.80 on the LOL dataset,23.31 dB and 0.87 on the MIT-Adobe FiveK dataset,and the target detection accuracy of the enhanced image on the ExDark dataset has increased by 7.6%,thus effectively improving the visualization quality of the low-illumination image and the target detection effect.
作者
杨伟
王帅
吴佳奇
陈伟
田子建
YANG Wei;WANG Shuai;WU Jiaqi;CHEN Wei;TIAN Zijian(School of Artificial Intelligence,China University of Mining and Technology(Beijing),Beijing 100083,China;Inner Mongolia Bureau of State Mine Safety Supervision Bureau,Hohhot 010010,China;School of Computer Science and Technology,China University of Mining and Technology,Xuzhou,Jiangsu 221116,China)
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2024年第4期158-169,共12页
Journal of Xi'an Jiaotong University
基金
国家自然科学基金资助项目(52274160)。
关键词
图像处理
机器视觉
轻量级网络
低照度图像
图像增强
目标检测
image processing
machine vision
lightweight network
low-light images
image enhancement
object detection
作者简介
杨伟(1995-),女,博士生;通信作者:田子建,男,教授,博士生导师。
