【作者】 张宇；

【导师】 蒋大钢；

【作者基本信息】 电子科技大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 无人机航拍是执行森林消防、农林植保、应急抢险等任务的基础。由于无人机飞行高度低、成像探测范围有限,应用时需通过图像拼接来扩大成像探测范围。但是,在图像特征稀少场景或时间敏感场景下,常用的基于图像特征的配准方法难以有效工作、基于灰度信息的配准方法不具有处理无人机多视角旋转仿射等变换的能力、基于位姿信息的配准方法受传感器精度限制配准精度较低。另外,在配准完成之后,需要保障全局(尤其是在重叠接缝处)色彩自然过渡,特别是在雾霾天气或因视角变化导致图像亮度差异较大的场景下,无论是采用等直方图规定化、wallis滤波等匀色处理方法进行全局色彩修正,还是采用直接平均融合、渐入渐出融合、最佳缝合线融合等融合方法,都难以实现色彩自然过渡。针对上述问题,本文提出了一种基于多灰度相似性测度表决的图像配准新方法和一种基于图像退化模型的拼接图像质量优化方法。具体研究工作概述如下:1)采用灰度相似性测度表决,实现图像特征稀少场景或时间敏感场景下的可靠快速配准。首先通过无人机位姿信息快速建立参考图像上部分像素点在待拼接图像上的初始映射,再利用差方和、互信息、直方图相关系数这三个灰度相似性测度进行表决修正映射点的位置,进而通过修正映射点求解单应变换矩阵参数,实现图像准确配准。实验结果表明:基于灰度相似性测度表决的配准方法能够在图像特征稀少场景下有效工作,耗时仅是基于SIFT特征提取的图像配准方法的15%左右,而且配准精度优于基于灰度信息和基于位姿信息的配准方法。2)基于图像退化模型修正透射率差异,保障亮度差异场景或雾霾天气下拼接图像接缝处的亮度自然过渡。首先采用暗通道先验算法进行大气光估计,结合图像退化模型,推导图像亮度与大气光、大气透射率的关系,再对存在色彩差异的图像进行亮度均衡或对雾霾图像进行去雾增强,从而生成全景图。实验结果表明:相对直方图规定化、wallis滤波、直接平均融合、渐入渐出融合、最佳缝合线融合等方法,通过透过率修正后的拼接图像在浓雾和色彩差异场景下的灰度均值差、标准偏差之差更小。3)将新的配准方法和拼接图像质量优化方法应用于不同飞行高度、不同视角的无人机集群图像拼接场景。实验结果表明:要发挥无人机集群快速探测优势,就需要减少不同机位的探测视场重叠区域,新的配准方法在重叠区域图像特征稀少的场景下更可靠,克服了无人机集群多机位成像的亮度差异。综上所述,本文提出的配准方法和拼接图像质量优化方法具有准确性、鲁棒性和快速性的综合优势,可在图像特征不明显、时间敏感场景、雾霾场景、集群探测场景下有效工作,为无人机或无人机集群在森林消防、农林检测、应急防护等领域中的应用提供新的航拍图像处理技术支撑。更多还原

【Abstract】 Unmanned aerial vehicle(UAV)aerial photography is the basis of forest fire prevention,agriculture and forestry plant protection,emergency rescue and other tasks.Due to the low flight altitude and the limited imaging detection range of the UAV,the application needs to expand the imaging detection range through image mosaicking.However,the commonly used registration method based on image features is difficult to work effectively in characterless scene or time-sensitive scene.The registration method based on grayscale information does not have the ability to handle the transformation such as multi-perspective rotational affine of UAV,and the registration method based on position and pose information has a large registration error under the limitation of sensor accuracy.In addition,the global(especially in the overlapping joints)color natural transition is need after registration.However,the haze weather or image brightness varies greatly due to the change of perspective,it is difficult to realize the global color correction,such as histogram specification and wallis filtering,or the fusion methods such as direct average fusion,gradual in and out fusion,and optimal suture fusion.According to the above problems,a new method of image registration based on multiple gray-scale similarity measure voting and a mosaic image quality optimization method based on image degradation model are proposed.The specific study work is summarized as follows:1)Gray-scale similarity measure voting is adopted to realize fast and reliable registration in characterless scene or time-sensitive scene.First of all,the initial mapping of the partial pixels between the reference image and the mosaic image are quickly established by the position and pose information of the UAV.Secondly,the position of the mapping points are corrected by voting on the three graysacle similarity measures:sum of differences,mutual information and histogram correlation coefficient.Finally,the modified mapping points are used to solve the homography to achieve accurate image registration.The experimental results show that the registration method based on the gray-scale similarity measure voting can work effectively in the scene with few image features,and the time consumption is only about 15%of the image registration method based on SIFT feature extraction,and the registration accuracy is better than the registration method based on gray information and the registration method based on position and pose information.2)Correcting the transmittance difference based on the image degradation model to ensure the natural brightness transition at the joint of the mosaic image under the brightness difference scene or haze weather.The dark channel prior algorithm is used to estimate atmospheric light.Combined with the image degradation model,to derive the relationship between image brightness and atmospheric light and atmospheric transmission.Then,the brightness of the image with color differences is balanced or the haze image is enhanced to generate a panoramic mosaic image.The experimental results show that compared to methods,such as histogram normalization,wallis filtering,direct average fusion,gradual fusion,and optimal suture fusion,the images corrected for transmittance have smaller mean difference and standard deviation in dense fog and color difference scene.3)The new registration method and mosaic image quality optimization method are applied to drone swarm images with different flight atitudes and perspectives.The experimental results show that in order to exploit the advantage of rapid detection of drone swarm,it is necessary to reduce the overlapping fields of different positions.The new registration method is more reliable in the scene with rare image features,and overcome the brightness difference of drone swarm.In summary,the registration method and the mosaic image quality optimization method have the comprehensive advantages of accuracy,robustness and speed,which can work effectively in characterless scene,time sensitive scene,haze scene,swarm detection scene,for UAV or drone swarm in forest fire protection,agriculture and forestry,emergency protection and other applications to provide new aerial image processing technology support.更多还原