DeepComp: A Hybrid Framework for Data Compression Using Attention Coupled Autoencoder

Due to the evolution of new media formats, emphasis on appropriate compression of data becomes paramount. Compression algorithms employed in real-time streaming applications must provide high compression ratio with acceptable loss. For such applications, the compression ratio of traditional compression algorithms used in Windows remains a challenge. Integrating deep learning algorithms with traditional Windows archivers can help the research objective in overcoming the challenges encountered by traditional Windows archivers. In this study, we propose a hybrid and robust compression framework named DeepComp that employs an attention-based autoencoder along with traditional Windows WinRAR archiver to compress both numerical and image data formats. Autoencoders- a well-known deep learning architecture widely used for data compression, outperform traditional archivers in terms of compression ratio but fall short in terms of reconstruction error. To minimize the reconstruction error, an attention layer is proposed in the autoencoder used in DeepComp. The attention layer accomplishes this by impeding the transition of spatial locality of the input data points during its processing in the compression and decompression phase. DeepComp is evaluated using numerical and image-type atmospheric and oceanic data obtained from the National Centers for Environmental Prediction (NCEP), which operates under National Oceanic and Atmospheric Administration (NOAA), USA. The performance analysis illustrates the robustness of DeepComp in compressing both numeric and image datatypes. In terms of compression ratio, it outperforms Windows archivers by an average of 69% and multilayer autoencoders by 48%. DeepComp also outperforms the reconstruction performance of the multilayer autoencoder.