一种曲线数据压缩的自编码器神经网络方法

Autoencoder neural network method for curve data compression

矢量空间数据的压缩是减少存储空间及节省网络传输带宽最直接有效的途径。本文利用自编码器神经网络构建了一种面向矢量曲线数据的压缩和解压模型。该模型针对不同规模和复杂度的曲线数据,首先进行分段和重采样处理进而规范化输入向量,然后通过优化自编码器结构和调整损失函数实现了高精度的压缩解码模块,最后通过弧段数据的坐标闭合差分配确保了数据的稳定性。对山西、湖南、江西3省的1∶100万县级弧段进行了压缩和还原试验,分析发现:数据还原精度随压缩率变小而变小,当压缩率小于35%后,数据还原精度呈波动趋势,25%为满足精度要求的适宜压缩率。比较傅里叶级数及贝塞尔曲线拟合方法发现本文模型压缩精度和处理速度在一定的压缩率范围内存在优势,同时该模型说明了深度学习在提取空间要素几何特征方面有一定的潜力。

Vector spatial data compression stands as the most direct and effective means for reducing storage space and conserving network transmission bandwidth.This study introduces a compression and decompression model tailored for vector curve data,leveraging autoencoder neural networks.Confronting curves of varying data volume and complexities,the model initiates with segmenting,resampling processes and normalizing input vectors.Through the optimization of autoencoder structure and adjustment of the loss function,a high-precision compression-decoding module is achieved.The stability of the data is ensured through closed coordinate differences of arc segment data.Compression and restoration experiments were carried out on the 1∶1000000 county-level arc segments in Shanxi,Hunan,and Jiangxi provinces.The analysis reveals that the accuracy of data restoration decreases with decrease in the compression rate.When the compression rate falls below 35%,the accuracy of data restoration exhibits a fluctuating trend.Therefore,a compression rate of 25% is recommended to ensure the required level of accuracy.Comparison with Fourier series and Bézier curve fitting methods indicates advantages in compression accuracy and processing speed within specific compression rate ranges for the proposed model.Additionally,the model highlights the potential of deep learning in extracting geometric features of spatial elements.In summary,this research presents a model for vector spatial data compression based on autoencoder neural networks,demonstrating promising performance and emphasizing the potential of deep learning in spatial feature extraction.