基于改进深度学习自编码的图像边沿畸变校正算法研究  

Image edge distortion correction algorithm based on improved deep learning self coding

作　　者：吴天强[1] 王义贺 WU Tian-qiang;WANG Yi-he(Taizhou Vocational College of Science&Technology,Taizhou Zhejiang 318020,China;Northeastern University,Shenyang Liaoning 110819,China;State Grid Liaoning Electric Power Company Limited Economic Research Institute,Shenyang Liaoning,110015,China)

机构地区：[1]台州科技职业学院,浙江台州318020 [2]东北大学,辽宁沈阳110819 [3]国网辽宁省电力有限公司经济技术研究院,辽宁沈阳110015

出　　处：《光电子．激光》2021年第2期149-156,共8页Journal of Optoelectronics·Laser

基　　金：国网辽宁省电力有限公司2018年第五批科技项目(第1项)-《能源互联网的时空异构模型构建与特性分析》(2018ZX-18)资助项目。

摘　　要：图像边沿畸变校正中直线投影衍生边沿拟合度差,导致校正误差大,提出基于改进深度学习自编码的图像边沿畸变校正算法。采用自适应阈值小波去噪算法,对各级尺度参数实施自适应变换,完成噪声去除。根据图像去噪结果,使用费舍尔向量编码优化深度学习结果,提取图像的边沿畸变形态。并以边沿畸变形态提取结果为基础,获取校正目标优化函数,分析边沿断裂情况,实现直线投影衍生边沿拟合;通过确定图像边沿误差评价函数,判断图像边沿畸变校正方式,达到图像边沿畸变校正的目的。以含噪桶形畸变与枕形畸变图像为研究对象进行实验分析,结果表明,所提算法可有效校正图像的桶形畸变与枕形畸变,桶形、枕形图像边沿畸变校正后,图像中的边沿条数和原图一致均为5条,实现高精度、高效率的图像边沿畸变校正。In the image edge distortion correction,the straight-line projection-derived edge fit is poor,resulting in large correction errors.An image edge distortion correction algorithm based on improved deep learning self-encoding is proposed.The adaptive threshold wavelet denoising algorithm is adopted to implement adaptive transformation of various scale parameters to complete noise removal.According to the results of image denoising,Fisher vector coding is used to optimize the deep learning results,and the edge distortion shape of the image is extracted.And based on the edge distortion shape extraction results,the correction objective optimization function is obtained,the edge fracture is analyzed,and the linear projection derivative edge fitting is realized;by determining the image edge error evaluation function,the image edge distortion correction method is judged to achieve the image edge distortion correction purpose.The experimental analysis is carried out on noisy barrel distortion and pincushion distortion images.The results show that the proposed algorithm can effectively correct the barrel and pincushion distortion of the image.After the edge distortion of the barrel and pincushion image is corrected,the image The number of edges is the same as 5 in the original image,realizing high-precision and high-efficiency image edge distortion correction.

关 键 词：改进深度学习自编码 图像畸变 衍生边沿 直线投影 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]