基于多模态深度学习的特定虚拟图像视觉特征自动补偿  

Automatic Compensation of Visual Features of Specific Virtual Images Based on Multimode Depth Learning

作　　者：何伟 杨大伟 马天福 马崇瑞 苑学贺 HE Wei;YANG Dawei;MA Tianfu;MA Chongrui;YUAN Xuehe(State Grid Xinjiang Information and Telecommunication Company,Urumqi,Xinjiang 830002,China;Beijing China-power Information Technology Co.,Ltd.,Beijing 100000,China)

机构地区：[1]国网新疆电力有限公司信息通信公司,新疆乌鲁木齐830002 [2]北京中电普华信息技术有限公司,北京100000

出　　处：《计算技术与自动化》2024年第3期102-107,共6页Computing Technology and Automation

基　　金：新疆维吾尔自治区科技攻关专项项目(2021B01013-6)。

摘　　要：为了提高特定虚拟图像视觉特征缺失补偿效果,提出了基于多模态深度学习的特定虚拟图像视觉特征自动补偿方法。首先,通过结合主成分分析(PCA)降维方法和结构张量的非局域全变分方法,对图像实行冗余信息去除和去噪处理;其次,通过稀疏自编码器和卷积池化技术展开深度学习,获取多模态下图像的彩色图、深度图、灰度图和3D曲面法线特征;最后,通过卷积神经网络完成特定虚拟图像视觉特征自动补偿。测试结果表明:这种特征补偿方法能够通过预处理提高图像清晰度,且图像视觉特征自动补偿质量较高。In order to improve the compensation effect for missing visual features of specific virtual images,a multimodal deep learning based automatic compensation method for visual features of specific virtual images is proposed.Firstly,by combining principal component analysis(PCA)dimensionality reduction method with structural tensor non local total variation method,redundant information is removed and denoised from the image.Secondly,deep learning is carried out through sparse autoencoder and convolutional pooling techniques to obtain color,depth,grayscale,and 3D surface normal features of images under multimodal conditions.Finally,automatic compensation of specific virtual image visual features is achieved through convolutional neural networks.The test results show that this feature compensation method can improve image clarity through preprocessing,and the quality of image visual feature automatic compensation is high.

关 键 词：多模态 特定虚拟图像 视觉特征自动补偿 图像去噪 卷积神经网络 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]