基于渐进式特征推理的裂缝图像修复方法研究  被引量:1

Research on crack image inpainting method based on progressive feature reasoning

在线阅读下载全文

作  者:李良福[1] 黎光耀 王楠[1] 张晰 LI Liangfu;LI Guangyao;WANG Nan;ZHANG Xi(College of Computer and Science,Shaanxi Normal University,Xi’an,Shaanxi 710000,China)

机构地区:[1]陕西师范大学计算机科学学院,陕西西安710000

出  处:《光电子.激光》2022年第4期393-402,共10页Journal of Optoelectronics·Laser

基  金:国家自然科学基金(61573232,61401263);陕西省自然科学基金(2022JM335)资助项目。

摘  要:图像修复是计算机图形学与计算机视觉中的研究热点之一。针对传统裂缝图像修复采用一次性补全修复方法并没有语义理解能力,当语义场景较为复杂且图像缺陷较大时修复效果不佳的问题,提出了一种基于渐进式特征推理的裂缝图像修复方法。该方法从孔洞边缘逐步修复图像,加强对孔洞中心的约束。首先通过计算掩膜占比确定修复比例,使用部分卷积更新掩膜,然后利用VGG-16网络进行特征提取,再使用语义注意力机制,生成高质量的图像特征,并采用跳跃连接方法,增强遥远距离的梯度相关性,为后续图像修复提供多尺度多层次的特征信息。最后,将递归生成的特征图进行融合解码生成修复图像。实验结果表明,与传统图像修复方法相比,本方法修复的裂缝图像峰值信噪比提升了0.5 dB—1.2 dB,产生语义明确的修复结果。Image inpainting is one of an activate research topic in the domain of computer vision and computer graphics.Aiming at the problem that the traditional crack image restoration method using one-time completion restoration method does not have the ability to understand semantics, and the repair effect is not good when the semantic scene is more complex and the image defect is large, a crack image restoration based on progressive feature reasoning is proposed.This method gradually restores the image from the hole edge and strengthens the constraint on the hole center.At first, partial convolution is used to update the mask, and the update ratio is determined by calculating the mask proportion.Then, use the VGG-16 network for feature extraction, semantic attention mechanism is used to generate high-quality image features, and use the jump connection method to enhance the gradient correlation of remote distances, so as to provide multi-scale and multilevel feature information for subsequent image restoration.Finally, the recursive feature map is fused and decoded to generate a repair image.The experimental results show that the proposed method, compared with traditional image inpainting methods, can improved the peak signal-to-noise ratio of the crack image repaired for 0.5 dB—1.2 dB and produce semantic clear inpainting results.

关 键 词:图像修复 注意力机制 部分卷积 裂缝图像 

分 类 号:TP391.9[自动化与计算机技术—计算机应用技术]

 

参考文献:

正在载入数据...

 

二级参考文献:

正在载入数据...

 

耦合文献:

正在载入数据...

 

引证文献:

正在载入数据...

 

二级引证文献:

正在载入数据...

 

同被引文献:

正在载入数据...

 

相关期刊文献:

正在载入数据...

相关的主题
相关的作者对象
相关的机构对象