光学学报, 2020, 40 (18): 1810002, 网络出版: 2020-08-27
基于卷积神经网络去噪正则化的条纹图修复 
下载： 1141次
Fringe Pattern Inpainting Based on Convolutional Neural Network Denoising Regularization
图 & 表
图 3. 不同方法确定的高光区域对比。(a)正常曝光的条纹图;(b)短曝光的条纹图; (c)图3 (a)的调制度;(d)图3 (b)的调制度;(e)Otsu方法对图3 (a)的结果; (f) Otsu方法对图3 (d)的结果
Fig. 3. Comparison of highlight areas determined by different methods. (a) Normal exposure time fringe pattern; (b) short exposure time fringe pattern; (c) modulation of Fig. 3 (a); (d) modulation of Fig. 3 (b); (e) result image of Fig. 3 (a) using Otsu threshold method; (f) result image of Fig. 3 (b) using Otsu threshold method
图 6. 修复结果。(a) 标准条纹图;(b)具有高光区域的模拟条纹图; (c)添加高斯噪声的迭代初值;(d)文献[ 19]方法的修复结果;(e)文献[ 20]方法修复结果;(f)本文方法修复结果
Fig. 6. Results of inpainting. (a) Standard fringe pattern; (b) simulated fringe pattern with highlight region; (c) initial value of the iteration with Gaussian noise; (d) inpainting result of Ref. [19] method; (e) inpainting result of Ref. [20] method; (f) inpainting result of proposed method
图 8. 不同初值的修复结果对比。(a)正常曝光条纹图;(b)直接对图8 (a)修复的结果;(c) 两帧融合得到的迭代初始图;(d)利用图8 (c)修复的结果
Fig. 8. Comparison of inpainting results with different initial values. (a) Normal exposure time fringe pattern; (b) inpainting result of Fig. 8 (a); (c) initial image-fused by proposed method; (d) inpainting result of Fig. 8 (c)
图 9. 条纹图修复结果。(a)原始条纹图;(b)迭代初值图;(c)文献[ 4]方法修复的结果;(d)文献[ 19]方法修复的结果;(e)文献[ 20]方法修复的结果;(f)本文方法修复的结果
Fig. 9. Fringe pattern inpainting results. (a) Original fringe pattern; (b) initial value for iteration; (c) inpainting result of Ref. [4] method; (d) inpainting result of Ref. [19] method; (e) inpainting result of Ref. [20] method; (f) inpainting result of proposed method
图 10. 相位恢复结果。(a) 文献[ 4]方法结果;(b)文献[ 19]方法结果;(c) 文献[ 20]方法结果;(d)直接用迭代初值的结果;(e)本文方法结果
Fig. 10. Phase reconstruction results. (a) Result of Ref. [4] method; (b) result of Ref. [19] method; (c) result of Ref. [20] method; (d) result of iterative initial value; (e) result of proposed method
图 11. 不同曝光时间下条纹图的灰度分布对比。(a)正常曝光时间条纹图;(b)短曝光时间条纹图;(c)本文方法修复的条纹图;(d)正常曝光时间条纹图和(e)修复过后条纹图第256行,170~370列的灰度分布
Fig. 11. Comparsion of gray distribution of fringe pattern under different exposure time. (a) Normal exposure time fringe pattern; (b) short exposure time fringe pattern; (c) inpainting result of proposed method; gray distribution of 170--370 columns in 256 row of (d) normal exposure time fringe pattern and (e) inpainting fringe pattern
表 1文献[ 19]方法、文献[ 20]方法和本文方法在用时、PSNR和重建相位的RMSE对比
Table1. Comparison in execution time, PSNR, and RMSE of phase reconstruction with Ref. [19] method, Ref. [20] method, and proposed method
|
彭广泽, 陈文静. 基于卷积神经网络去噪正则化的条纹图修复[J]. 光学学报, 2020, 40(18): 1810002. Guangze Peng, Wenjing Chen. Fringe Pattern Inpainting Based on Convolutional Neural Network Denoising Regularization[J]. Acta Optica Sinica, 2020, 40(18): 1810002.
PDF全文
