基于改进双向注意力映射的单板图像修复  

Image Inpainting of Veneer with Improved Learnable Bidirectional Attention Maps

作　　者：娄蕴祎 张冬妍[1] 葛奕麟 崔明迪 张泽冰 LOU Yunyi;ZHANG Dongyan;GE Yilin;CUI Mingdi;ZHANG Zebing(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学机电工程学院,哈尔滨150040

出　　处：《森林工程》2023年第2期132-138,共7页Forest Engineering

基　　金：林业公益性行业科研专项(201504307)。

摘　　要：木材生长加工过程中产生的缺陷会影响产品质量并且浪费大量木材资源,为提高木材利用率与缺陷修复效果,提出一种基于可学习的双向注意力映射(Learnable Bidirectional Attention Maps,LBAM)网络模型的轻量化Lightweight LBAM网络(LL-Net)。该网络使用级联与并行方式的膨胀卷积扩大感受野,修改掩膜更新的激活函数提高修复效果,减少网络深度,在保证效果前提下降低参数量。结果表明,LL-Net与全局与局部判别器(Global and Local Discriminator,GL)方法相比,峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)和结构相似性(Structural Similarity,SSIM)最高分别提升48.6%和14.2%;与上下文注意力(Contextual Attention,CA)方法相比,PSNR和SSIM最高分别提升23.0%和7.9%;与LBAM方法相比,PSNR和SSIM最高分别提升1.5%和0.6%。并且LL-Net网络参数量为63.58 m,相较于LBAM方法降低了75%。该方法可取得纹理更清晰、语义一致性更好的修复效果,为单板缺陷修复提供指导性意见。Defects in the growth and processing of wood will affect product quality and waste a significant amount of wood.To improve the use rate of wood and defect repair effect,this study proposes a Lightweight LBAM Network(LL-Net)for veneer based on learnable bidirectional attention maps(LBAM).In this network,cascade and parallel dilated convolution were utilized to increase the receptive field,and activation function of mask update was modified to improve the repair effect,reduce the network depth,and reduce the number of parameters on the premise of ensuring the effect.The results showed that compared with Global and Local Discriminator(GL)method,the Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity(SSIM)were increased by 48.6%and 14.2%,respectively.Compared with Contextual Attention(CA),the PSNR and SSIM were increased by 23.0%and 7.9%,respectively.Compared with LBAM,the PSNR and SSIM were increased by 1.5%and 0.6%,respectively.The number of LL-Net network parameters was 63.58 m,which was 75%lower than that of LBAM method.This method can achieve clearer texture and better semantic consistency,and provide guidance for veneer defect repair.

关 键 词：图像修复 深度学习 单板 双向注意力映射 

分 类 号：S781.5[农业科学—木材科学与技术] TP391.4[农业科学—林学]