基于深度学习U-net网络的雾天汽车视觉图像超像素级配准方法  

Super pixel level registration method for foggy car vision images based on deep learning U-net network

作　　者：靳新[1] 潘月 JIN Xin;PAN Yue(Shenyang Institute of Technology,Shenyang 113122,China)

机构地区：[1]沈阳工学院,沈阳113122

出　　处：《激光杂志》2025年第4期121-127,共7页Laser Journal

基　　金：辽宁省教育科学“十四五”规划2022年度课题(No.JG22DBF514)。

摘　　要：雾天汽车视觉图像因对比度降低和细节模糊而难以处理与配准。为此,提出基于深度学习U-net网络的超像素级配准方法。首先,通过改进的直方图均衡化算法,增强雾天图像的清晰度和对比度。接着,利用结合了GAN技术的U-Net网络对图像进行初始分割,获取不同区域的标签集。随后,应用SLIC超像素分割算法,将相似像素组合成超像素,形成更具代表性的图像特征。最后,采用改进SURF算法,利用超像素特征进行精确图像对齐,提高配准精度和效率。实验证明,此方法不仅能有效改善雾天汽车视觉图像质量,还具备高配准精度,NCC值稳定在0.92至0.95之间。The visual images of foggy cars are difficult to process and register due to reduced contrast and blurred details.To this end,a superpixel level registration method based on deep learning U-net network is proposed.Firstly,by improving the histogram equalization algorithm,the clarity and contrast of foggy images are enhanced.Next,the U-Net network combined with GAN technology is used to perform initial segmentation of the image and obtain label sets for different regions.Subsequently,the SLIC superpixel segmentation algorithm is applied to combine similar pixels into superpixels,forming more representative image features.Finally,an improved SURF algorithm is adopted to accurately align images using superpixel features,improving registration accuracy and efficiency.Experimental results have shown that this method not only effectively improves the visual image quality of foggy vehicles,but also has high registration accuracy,with NCC values stable between 0.92 and 0.95.

关 键 词：直方图均衡化 深度学习GAN-U-net分割网络 SLIC超像素分割 SURF超像素级配准 

分 类 号：TN911[电子电信—通信与信息系统]