Research On Sequential And Spatial Data Compression Technology Based On Deep Learning

With the development of computer technology,human has entered the information age,and the total amount of data flowing in the computer world has increased rapidly.The storage and effective transmission of data are huge challenges.Data compression can alleviate the difficulties in data storage and transmission.In recent years,thanks to the development of deep learning technology,the field of data compression has made great progress.This research explores the application of deep learning methods in some forms data compression task and achieve leading performance.Real world data has different forms,and two types are representative:(1)One-dimensional serialized data,such as audio,text,genetic data,etc.This type of data presents the autocorrelation before and after the sequence,and exists time redundancy and entropy redundancy;(2)Two-dimensional spatial data,such as natural images,satellite remote sensing images,medical images,etc.,this kind of data presents autocorrelation in spatial position,and exists spatial redundancy,structural redundancy,perception redundancy and entropy redundancy.This research explores the compression methods of one-dimensional serialized data and two-dimensional spatial data respectively.We take gene quality values and natural images as the research objects,and choose reasonable compression strategies according to their compression requirements.The elements in the gene quality values characterize the measurement error rate of sequencing sites,and play an important role in downstream analysis tasks such as mutation site detection.This work compress quality values lossless.Traditional image compression algorithms take visual effects and coding efficiency as their optimization goals,while deep learning image compression algorithms mostly take reconstruction errors and coding efficiency as their optimization goals.Nowadays,images are widely used in image classification tasks.However,the reconstructed images of traditional image compression methods or deep learning based image compression methods will decrease the accuracy of image classification models.This research takes image applications in classification tasks into account,and studies image compression algorithms driven by classification tasks.In summary,this research mainly studies two works:(1)This research proposes a deep learning based lossless compression method for gene quality values.The algorithm includes a gene quality values prediction model based on the LSTM network and an arithmetic encoder.The prediction model makes full use of the proximity correlation and position correlation of the quality values,and can efficiently learn and predict the conditional probability distribution of the quality values.This word achieves leading compression performance on gene quality values.(2)This research proposes an image compression model training framework driven by image classification task.This training framework is aims to improve the classification performance of reconstructed images.Experiments show that the reconstructed image of the image compression model retrained by this framework has a higher classification accuracy under the same bit rate.