Research On Visual Data Compressed Sensing Algorithm Based On Deep Learning

With the rapid development of information technology,there is a growing demand for real-time processing of massive visual data.How to quickly realize data uplink transmission and downlink response tasks is particularly important.Compressed Sensing(CS),as a sampling theory based on signal sparsity,allows for sampling ratios significantly lower than the Nyquist-Shannon sampling theorem,enabling fast and efficient signal recovery.The simultaneous sampling and compression characteristics of CS transfer the computational complexity to the decoding end,enabling simplicity and efficiency on the encoding end.Its swift compression and high-quality reconstruction performance are instrumental for facilitating real-time transmission and efficient analysis and response in IoV applications,particularly in urban safety and autonomous driving domains.Additionally,the introduction of CS provids a new approach and method for data processing in resource-constrained environments,propelling the further application of vehicular networking in areas such as urban safety and autonomous driving,and accelerating the deployment of vehicular networking in-car terminals.Based on the aforementioned research background,this paper addresses the practical needs of fast transmission,real-time response,noise resistance,and lightweight hardware deployment for vision data.It combines CS technology with powerful and efficient deep learning technology,and proposes the image/video CS algorithm models with faster speed,better performance and lighter structure.Specifically,this study mainly includes the following four aspects:Firstly,To solve the problem that the current CS algorithm does not fully utilize the sparsity of the image and the complexity of the reconstruction algorithm is high,two novel image compressed sensing algorithms based on wavelet transform are proposed,denoted as WCS-Nets(WCS-Net and WCS-Net+).Among them,the WCS-Net algorithm is based on the DR2-Net algorithm and combines the sym8 wavelet transform with the sampling network to fully exploit the low redundancy characteristics of the image and achieve the acquisition of high-fidelity measurement values.The WCS-Net+algorithm uses a multi-scale residual learning network to better optimize the reconstruction quality of the image,constituting an enhanced version of the WCS-Net algorithm.Experimental results demonstrate that our proposed methods outperform state-of-the-art CS algorithms in terms of reconstruction quality,runtime efficiency,and robustness to noise.Secondly,In view of the problem that the above algorithm does not explore the impact of wavelet transform series selection and sampling and reconstruction network design on the quality of the compressed sensing algorithm,a hierarchical deep network image compressed sensing algorithm based on multi-level wavelets is proposed(MWHCSNet).MWHCS-Net consists of three modules:a sampling module based on a multilevel wavelet transform,a hierarchical initial reconstruction module and a lightweight deep reconstruction module.The sampling module based on multilevel wavelet transform with hierarchical subspace learning for progressive acquisition of measurements to further optimize sampling efficiency and stability.To enhance the finer texture details,the hierarchical initial reconstruction module is designed as a basic initial reconstruction network plus an enhanced initial reconstruction network,which corresponding to the dominant structure component and the texture detail component of the reconstructed image,respectively.Experimental results demonstrate that MWHCS-Net achieves superior reconstruction performance while maintaining comparable runtime.Moreover,MWHCS-Net exhibits better robustness against noise in most cases compared to the existing deep learning-based image CS methods.Thirdly,to further explore the construction of lightweight sampling matrices,a scalable compressive sampling network based on progressive hierarchical subspace learning is proposed(SPHSL-CSNet).SPHSL-CSNet is is jointly trained by using progressive hierarchical subspace learning in an end-to-end manner.Specifically,three wavelet levels are employed for image sparse transformation,achieving band separation via frequency band sensing masks.Independent sampling is performed on specific frequency bands(low-frequency,low-mid-frequency,low-mid-high frequency,and the entire wavelet band),effectively reducing the parameter count of the sampling matrix.Experiments show that the proposed algorithm obtains high-quality reconstruction with fewer network parameters and pays more attention to the restoration of texture details,which is more conducive to deployment on devices with limited computing resources or small storage space.At the same time,it also has a good effect on multi-spectral image compression.In addition,SPHSL-CSNet also performs well in terms of denoising performance.Lastly,Aiming at the problem that the current video CS algorithm does not make full use of the intra-frame and inter-frame correlation and sparsity of the video sequence,and the reconstruction network sacrifices storage space to pursue high-quality reconstruction,a new,lightweight video CS algorithm based on wavelet residual sampling and multi-domain fusion is proposed(WRMD).Specifically,the wavelet residual sampling selects the keyframe near the nonkeyframe position,constructs the wavelet domain residual between it and the nonkeyframe,and realizes efficient and high-quality measurements adaptively.Moreover,To enhance the reconstruction quality of nonkeyframes,the stradgy of multi-domain fusion is adopted,which fuses the information of the pixel domain,dual keyframe multi-level feature domain and residual frame multi-level feature domain for optimization.Extensive experiments prove that WRMD surpasses the the state-of-the-art video CS methods and deep learning-based image CS methods in both subjective and objective evaluations,and has excellent algorithm performance.