基于卷积网络模型的低照度激光图像特征增强方法  

Feature enhancement method of low illumination laser image based on deep learning

作　　者：周恺[1,2] 李婧 ZHOU Kai;LI Jing(School ofInformation Engineering,Nanjing Xiaozhuang University,Nanjing 211171,China;Key Laboratory of Intelligent Information Processing,Nanjing 211171,China)

机构地区：[1]南京晓庄学院信息工程学院,南京211171 [2]南京市智能信息处理重点实验室,南京211171

出　　处：《激光杂志》2023年第12期120-125,共6页Laser Journal

基　　金：国家自然科学基金青年项目(No.62205150);江苏省自然科学基金青年项目(No.BK20210036);江苏省高等学校自然科学研究项目(No.21KJB510013)。

摘　　要：针对当前低照度激光图像特征增强方法在图像增强过程中需进行多次翻转平移处理,造成激光图像特征增强后损失值较大的问题,提出基于卷积网络模型的低照度激光图像特征增强方法。应用激光图像色彩模型以及去噪自编码器,完成低照度激光图像预处理。使用分段性变换方法设计激光图像映射关系函数,得到低照度激光图像增强目标函数。构建卷积神经网络模型以及模型对应损失函数,完成低照度激光图像特征增强。至此,基于卷积网络模型的低照度激光图像特征增强方法设计完成。实验结果表明:此方法增强真实与合成图像特征后的损失值较低,分别为0.245和0.361,其峰值信噪比较高,分别为45.52和48.54,极大地提高了图像的应用价值,且文中方法对图像增强处理的时长最短,在13 s到16 s之间,其应用性能较高。Aiming at the problem that the current low illumination laser image feature enhancement methods need to carry out multiple flipping and translation processing in the process of image enhancement,resulting in a large loss value after laser image feature enhancement,a low illumination laser image feature enhancement method based on con-volution network model is proposed.Using laser image color model and denoising self encoder,low illumination laser image preprocessing is completed.The mapping function of laser image is designed by using the segmented transforma-tion method,and the objective function of low illumination laser image enhancement is obtained.The convolution neu-ral network model and its corresponding loss function are constructed to complete the feature enhancement of low illu-mination laser image.So far,the design of low illumination laser image feature enhancement method based on convolu-tion network model has been completed.The experimental results show that the loss value of this method after enhan-cing the features of real and synthetic images is low,which is 0.245 and 0.361 respectively,and its peak signal to noise ratio is high,which is 45.52 and 48.54 respectively,which greatly improves the application value of images,and the method in this paper has the shortest processing time for image enhancement,which is between 13s and 16s.

关 键 词：卷积网络模型 图像滤波处理 图像增强 损失函数 低照度激光图像 处理时长 

分 类 号：TN911[电子电信—通信与信息系统]