基于多头注意力与双向门控循环单元的数据无损压缩方法
Lossless Data Compression Method Based on Multi-Head Attention and Bidirectional Gated Recurrent Unit

为解决数据归档存储场景中出现的物理存储成本增长和数据库内存紧张等问题，本文提出一种基于注意力机制与双向门控循环单元的数据无损压缩方法，采用Transformer和双向门控循环单元作为概率预测器，输出数据流的条件概率分布，结合自适应算术编码器对数据进行压缩。实验对比结果表明，本文所提方法相较于算术编码和基于字典模型的LZW这两种传统无损压缩方法，压缩率分别平均提升约28.8%和7.8%；相较于Cmix v19和NNCP两种深度学习方法，平均压缩率分别降低0.4%和0.2%，但平均压缩时间分别约为其5.1%和39.4%。
To solve the problems of rising physical storage cost and memory of database limitation in the archiving and storing scenario of data, a lossless data compression method based on multi-head attention and bidirectional gated recurrent unit is proposed in this paper. Transformer and bidirectional gated cycle unit are used as probability predictors to output conditional probability distribution of data flow and an adaptive arithmetic encoder is combined to compress the data. The exper-imental results show that compared with the two traditional lossless compression methods, arithmetic coding and LZW based on the dictionary model, the compression ratio of the proposed method is improved by 28.8% and 7.8% respectively. Compared with Cmix v19 and NNCP, the average compression rate is reduced by 0.4% and 0.2%, respectively, but the average compression time is about 5.1% and 39.4%, respectively.