一种鲁棒的指静脉近红外图像质量修复模型OA
A robust model for restoring near-infrared image quality of finger veins
针对现有图像修复方法多面向单一退化类型设计,难以在复杂退化环境下同时兼顾结构恢复与细节增强的问题,本研究在分析图像退化物理模型的基础上,提出一种可针对不同频率成分,进行自适应增强的图像质量修复模型.首先,利用傅里叶变换及可学习频率掩码,对退化图像进行高、低频分离;然后,针对图像的低频和高频成分,分别引入方向注意力和空间自适应特征调制模块,构建特征增强路径,并利用频率交叉注意力融合机制,提升修复性能.本研究所提模型在指静脉数据集 FV-USM 和 MMCBNU_6000 上的实验结果表明,该模型峰值信噪比(PSNR)分别达到 40.8218、42.5337 dB,结构相似性指数(SSIM)分别为 0.9842、0.9873.本研究模型在处理不同类型指静脉退化图像时,表现出优异的适应性和鲁棒性,提高了身份识别的准确率与稳定性.
In order to address the issue that the existing image inpainting methods are mostly designed for a single type of degrada-tion,making it difficult to simultaneously achieve both structural restoration and detail enhancement in complex degradation environ-ments,we proposed an adaptive image quality inpainting model for different frequency components based on the analysis of the physical model of image quality degradation.Firstly,the Fourier transform and its learnable frequency mask were used to separate the high and low-frequency components of the degraded image.Then,for the low-frequency and high-frequency components of the image,direc-tional attention and spatial adaptive feature modulation modules were introduced to build feature enhancement paths,and the frequency cross-attention fusion mechanism was used to improve the restoration performance.The proposed model was validated on the finger vein datasets FV-USM and MMCBNU_6000.Experimental results showed that the proposed model achieved peak signal to noise ratio(PSNR)values of 40.8218 and 42.5337 dB,and structural similarity index measure(SSIM)values of 0.9842 and 0.9873,respective-ly.This research demonstrates excellent adaptability and robustness when dealing with different types of degraded finger vein images,improving the accuracy and stability of identity recognition.
赵英梦;李涵;刘善奇;李彦蓉;刘庆一;白培瑞
山东科技大学 电子信息工程学院,青岛 266590山东科技大学 电子信息工程学院,青岛 266590山东科技大学 电子信息工程学院,青岛 266590山东科技大学 电子信息工程学院,青岛 266590山东科技大学 电子信息工程学院,青岛 266590山东科技大学 电子信息工程学院,青岛 266590
医药卫生
指静脉图像修复生物特征识别频率分离方向滤波交叉注意力
Finger vein image inpaintingBiometric recognitionFrequency separationDirectional filteringCross-attention
《生物医学工程研究》 2026 (4)
288-296,9
评论