基于神经网络的飞行器空中侦查时低照度图像增强系统  

Neural Network Based Low-Light Aerial Image Enhancement System for Aerial Surveillance by Aircraft

作　　者：杨允权 刘国宁[1] 孙克文 杨洁[1] YANG Yun-quan;LIU Guo-ning;SUN Ke-wen;YANG Jie(School of Mechanical and Power Engineering,Zhengzhou University,He’nan Zhengzhou 450001,China)

机构地区：[1]郑州大学机械与动力工程学院,河南郑州450001

出　　处：《机械设计与制造》2023年第9期163-168,共6页Machinery Design & Manufacture

基　　金：河南省产学研重点支持项目(172107000008)。

摘　　要：飞行器航拍时,由于飞行环境和飞行姿态的影响,会导致图像出现细节丢失和颜色失真从而影响图片质量,对图像中的目标准确识别带来挑战。针对这一问题,提出了一种Ritinex理论和基于神经网络深度学习相结合的方法,利用低照度图像增强算法来提高图像质量,从而实现侦查目标的准确识别。在AlexNet神经网络构架中引入U-net结构,根据Retinex理论采用分组卷积和深层卷积块构建了用于低照度图像增强神经网络构架,并加入亮度调整因子对分解结果进行增强。实验表明,本研究提出的方法和其它常用算法相比,有较好的整体视觉效果。另外,基于该方法所开发的低照度图像增强系统在搭建的无人机航拍平台上的应用也显示了其可靠性和准确性。The quality of aerial image is greatly affected by the flight pose of the aircraft and the surrounding environment which can lead to the loss of details and color distortion.And this is main challenge for accurate target identification in aerial image.In order to solve this problem,an approach with the combination of Ritinex theory and deep learning method based on neural net⁃works is proposed to enhance the low-light image quality with image enhancement algorithm and realize accurate target identifi⁃cation.The neural network consisting of grouped convolution blocks and deep convolution blocks is made by adopting the AlexNet architecture including U-net and considering the Retinex theory.Brightness enhancement factors are also included in the neural network.The experimental results show that the approach in this study has better overall visual effects.In addition,the developed low-light image enhancement system based on the proposed approach here is tested and practiced on the drone aerial platform to identify the targets and test results reveal its reliability and accuracy.

关 键 词：空中侦查 RETINEX理论 低照度 图像增强 深度学习 

分 类 号：TH16[机械工程—机械制造及自动化]