CNN图像修复区域故障检测的迭代方法

Iterative Application of Fault Detection Based on CNN Image Repair Area

传统的基于补丁的图像修复方法在修复过程中无法复制真实的图像纹理,即故障像素区域,通常需要手动标记故障区域.我们提出了一个迭代检测故障区域的图像修复模型.使标记故障区域过程自动化、精确化,模型将从修复过程中的多个特征图中图区的补丁作为CNN的输入,经CNN卷积和池化后输入Softmax分类器识别区域像素有效性,对标记为无效的故障区域采用Patch Match算法进行修复,引入启发式阈值对修复图像多次迭代直至收敛.实验结果表明,本文方法与传统的人工标记法相比节约了人工时间,与非迭代应用方法相比,降低了故障像素与原始缺失区域比值,证实了方法的有效性.

Traditional patch-based image restoration methods can’t reproduce the real image texture in the process of restoration,that is,the fault pixel area,which usually needs to be labeled manually.We propose an image restoration model for iteratively detecting the fault area.This model makes the process of labeling the fault area automated and accurate,and the model will be from the repair process.Patch Match algorithm is used to repair the invalid fault area.Heuristic threshold is introduced to iterate the repaired image several times until convergence.The experimental results show that the proposed method is effective.Compared with the traditional manual labeling method,the proposed method saves manual time and reduces the ratio of fault pixels to the original missing area compared with the non-iterative application method,which proves the effectiveness of the proposed method.