基于傅里叶变换和差分正则化的改进SimVP时空图像序列预测模型  

Improved SimVP Spatiotemporal Image Sequence Prediction Model Based on Fourier Transform and Differential Regularization

作　　者：冯耀祖 胡凯 Feng Yaozu;Hu Kai(College of Internet of Things Engineering,Jiangnan University,Wuxi 214122,Jiangsu,China;Key Laboratory of Light Industry Process,Jiangnan University,Wuxi 214122,Jiangsu,China)

机构地区：[1]江南大学物联网工程学院,江苏无锡214122 [2]江南大学轻工过程重点实验室,江苏无锡214122

出　　处：《科技通报》2025年第4期42-50,共9页Bulletin of Science and Technology

基　　金：国家自然科学基金项目(71904064)。

摘　　要：在时空图像序列预测领域中,为了提高预测精度,各类模型的复杂度逐渐增加。其中,SimVP (simpler yet better video prediction)另辟蹊径,在保持良好预测效果的同时,追求模型结构的简化。SimVP完全基于CNN (convolutional neural networks)构建,避免了使用复杂的RNN (recurrent neural network)、LSTM (long short-term memory)和Transformer等模块。然而,基于CNN的模型在处理长序列预测时可能存在局限性,此外其泛化能力尚未得到充分验证,特别是在处理更加复杂或多样化的数据时。因此,本文提出了一种基于SimVP的改进方案,在模型结构中增加了傅里叶变换模块,在频域提取图像数据以提高模型在处理无规则形状与弱变化规律数据时的效果。同时,在模型训练过程中引入了差分散度正则化项,以解决SimVP在处理长序列数据时帧间差异扩大的问题。为验证模型处理特殊数据的效果,使用气象数据检验模型预测不定型目标的能力,使用文献数据检验模型预测表现抽象数据的能力。实验结果显示:改进的模型在气象数据集上平均绝对误差(mean absolute error,MAE)下降0.268%,均方误差(mean-square error,MSE)下降0.580%,结构相似性指数(structural similarity,SSIM)提高0.464%,峰值信噪比(peak signal-to-noise ratio,PSNR)提高0.154%,在文献数据集上MAE下降4.619%。由结果可知改进后的模型在特定复杂数据的预测精度上有一定的提升,扩大了模型的泛用性。此外,差分散度模块通过减小帧间差异的方式提高了模型对序列数据的整体预测精度,在一定程度上弥补了模型在捕捉长距离时间依赖关系方面的缺陷。In the domain of spatiotemporal image sequence prediction,the complexity of various models has progressively escalated in pursuit of enhanced prediction accuracy.Among these,SimVP(simpler yet better video prediction) adopts a distinctive approach,striving for model structure simplification while preserving commendable prediction performance.SimVP is entirely constructed on CNN(convolutional neural networks),eschewing the utilization of intricate modules such as RNN(recurrent neural network),LSTM(long short-term memory),and Transformers.However,models based on CNNs may encounter limitations when addressing long-sequence predictions,and their generalization capabilities have not been fully substantiated,particularly when managing more complex or diverse data.Consequently,an enhancement scheme based on SimVP is proposed,which integrates a Fourier transform module into the model architecture to extract image data in the frequency domain,thereby augmenting the model's performance when processing data with irregular shapes and subtle change patterns.Simultaneously,a differential divergence regularization term is introduced during the model training phase to mitigate the issue of escalating frame differences when SimVP processes long-sequence data.To validate the model's proficiency in handling special data,meteorological data is utilized to assess the model's capacity to predict shapeless targets,and bibliographic data is employed to evaluate the model's ability to predict abstract data.The experimental results demonstrate that the improved model achieves a reduction of 0.268% in MAE(mean absolute error),a decrease of 0.580% in MSE(mean-square error),an augmentation of 0.464% in SSIM(structural similarity),and an enhancement of 0.154% in PSNR(peak signal-to-noise ratio) on the meteorological dataset,on the bibliographic dataset,the MAE diminishes by 4.619%.These findings indicate that the improved model has made notable advancements in prediction accuracy for specific complex data,thereby broadening the model's versati

关 键 词：图像序列预测 时空特征 SimVP模型 帧间差分 傅里叶变换 

分 类 号：TP391.7[自动化与计算机技术—计算机应用技术]