Research On Image Compression Method Based On Learnable Multi-frequency Representatio

Nowadays the information technology improves rapidly,social media sharing,autonomous driving,satellite communication,video surveillance and other application scenarios have a strong demand for image compression,and it becomes very important to obtain satisfactory compressed images based on limited hardware resources.As a new type of remote sensing image,hyperspectral image contains rich spatial information,spectral information,radiation information,etc.,which is widely used in resource census,environmental protection,geological detection and other fields.However,hyperspectral images generally have more spectral dimensions and a large amount of data,which brings challenges to data storage and transmission.Therefore,hyperspectral image compression is also very important.With the development of deep learning technology,new possibilities are emerging in the field of image compression.Image compression using deep learning can jointly optimize each module in the framework by learning a large number of image data,which is more efficient than manual module design.In addition,each module is implemented by neural network,and the whole framework can be jointly tuned by gradient back propagation.In view of this feature,this paper uses the deep learning framework of end-to-end optimization to complete the tasks of common image compression and hyperspectral image compression.The main tasks are as follows:1.An enhanced multi-frequency representation learned image compression method is proposed.Based on the existing multi-frequency learned image compression method based on generalized octave convolution,this paper proposes to add channel attention module to further reduce spatial redundancy and add decoder enhancement module to enhance the quality of recovered images.The algorithm showed excellent visual effects on standard test data set Kodak compared with other standard codecs and advanced learning-based image compression methods,and achieved better results on MS-SSIM metrics at low bit rates.2.An entropy-optimized multi-frequency representation learned image compression method is proposed.In this paper,based on the enhanced multi-frequency representation learned image compression method,the entropy model is improved.The probability distribution of the latent variables is updated as mixture gaussian distribution,from original single gaussian distribution.Besides,the global attention module combined with finer parameter estimation module to improve the potential probability of latent variables is added,for further optimize the compression ability.The improved method performs better in MSSSIM metrics,especially at 0.134 BPP and 0.243 BPP ratio.3.A hyperspectral image lossless compression method based on learned wavelet is proposed.The multi-resolution idea of Octave convolution is consistent with wavelet analysis,while the learning wavelet is more explanatory than the black box convolution.In this paper,3D convolutional neural networks are applied as predictors and renovators of lifting wavelet transforms,and a compression framework for hyperspectral images is trained by end-to-end learning.The proposed algorithm performs well in CAVE data sets,compared with JP3D compression method which is exposed by OpenJPEG2.4.0 and the compression ratio is around 2.5.Compared with the original 3D Le Gall 5/3 wavelet of JP3D,the results show that the learned wavelet has better compression performance.