基于深度学习的沉浸式投影系统图像反射补偿问题研究  

Research on Image Reflection Compensation of Immersive Projection System Based on Deep Learning

作　　者：马钰[1] MA Yu(Shaanxi Institute of International Trade&Commerce,Xi’an 712046,China)

机构地区：[1]陕西国际商贸学院,西安712046

出　　处：《自动化与仪器仪表》2022年第12期20-24,30,共6页Automation & Instrumentation

基　　金：陕西省教育厅2021年度一般专项科研计划项目《沉浸式艺术教育对当代大学生审美能力的培养研究》(21JK0047)。

摘　　要：针对传统补偿网络模型PCN生成图像边缘丢失、精度缺失和补偿效果不佳的问题,提出基于补偿网络模型PCN,分别引入超清机制和感知损失SR层,得到基于卷积神经网络的补偿网络优化模型SRCN。实验结果表明,将SRCN补偿网络模型应用到多投影系统数据集后,SRCN模型的SSIM、PSNR客观指标分别取值为0.93和20.51,相较于基础的PCN图像反射补偿模型,本模型的SSIM、PSNR值高出了0.25和2.18。且本模型的RMSE值为0.09,比PCN模型低了0.05。由此可知,本模型在PCN模型的基础上加入超清机制和感知损失后,生成的图像纹理细节更加丰富,图像的人眼感知度更加清晰,无图像边缘模糊问题,说明优化后的模型具备可行性。In view of the problems of image edge loss, accuracy loss and poor compensation effect of the traditional compensation network model PCN, this paper proposes the compensation network model PCN, introduces the superdefinition mechanism and perceptual loss SR layer respectively, and obtains the compensation network optimization model SRCN based on convolutional neural network. Experimental results show that, after the SRCN compensation network model is applied to the multi-projection system data set, the SSIM and PSNR objective indicators are 0.93 and 20.51, respectively. The SSIM and PSNR values are 0.25 and 2.18 higher than that compared to the basic PCN image reflection compensation model. Moreover, the RMSE value of this model is 0.09, which is 0.05 lower than the PCN model. Therefore, after this model adds the superdefinition mechanism and perception loss to the PCN model, the resulting image texture details are more rich, the human eye perception is clearer, and there is no image edge blur, which shows that the optimized model is feasible.

关 键 词：卷积神经网络 FCN 投影系统 图像反射补偿 SRCN 

分 类 号：TP392[自动化与计算机技术—计算机应用技术]