Deep Network for Image Compressed Sensing Coding Using Local Structural Sampling

TL;DR

This paper introduces CSCNet, a CNN-based image compressed sensing coding framework that integrates local structural sampling, measurement coding via a third-party codec, and a convolutional Laplacian pyramid for progressive reconstruction. A learnable, highly sparse local sampling matrix replaces traditional random matrices to produce highly correlated measurements, which are then efficiently encoded and decoded to recover the image. The model is trained end-to-end, balancing reconstruction quality and rate through a loss that includes a TV-based rate term and a reconstruction term. Experimental results show CSCNet consistently outperforms existing CSC and CSBC methods and offers competitive performance against traditional codecs, while delivering faster encoding and scalable multi-resolution reconstruction. These findings suggest CS-based coding with learnable local sensing can achieve superior rate-distortion trade-offs in resource-constrained imaging scenarios.

Abstract

Existing image compressed sensing (CS) coding frameworks usually solve an inverse problem based on measurement coding and optimization-based image reconstruction, which still exist the following two challenges: 1) The widely used random sampling matrix, such as the Gaussian Random Matrix (GRM), usually leads to low measurement coding efficiency. 2) The optimization-based reconstruction methods generally maintain a much higher computational complexity. In this paper, we propose a new CNN based image CS coding framework using local structural sampling (dubbed CSCNet) that includes three functional modules: local structural sampling, measurement coding and Laplacian pyramid reconstruction. In the proposed framework, instead of GRM, a new local structural sampling matrix is first developed, which is able to enhance the correlation between the measurements through a local perceptual sampling strategy. Besides, the designed local structural sampling matrix can be jointly optimized with the other functional modules during training process. After sampling, the measurements with high correlations are produced, which are then coded into final bitstreams by the third-party image codec. At last, a Laplacian pyramid reconstruction network is proposed to efficiently recover the target image from the measurement domain to the image domain. Extensive experimental results demonstrate that the proposed scheme outperforms the existing state-of-the-art CS coding methods, while maintaining fast computational speed.

Figures (9)

• Figure 1: Diagram of the proposed image CS coding framework CSCNet. Three functional modules are included: local structural sampling, measurement coding and Laplacian pyramid reconstruction.
• Figure 2: Diagram of the proposed local structural sampling. LO is the localization operation, and PNO is the positive normalization operation. The dashed box shows the configuration of our local sampling network.
• Figure 3: The details of the Laplacian pyramid reconstruction network (red arrow). The architectures in the dashed boxes show the details of "Transform+Concat" and the structure of each level respectively. The double values in brackets represent the channel number of input feature and output feature respectively.
• Figure 4: The thumbnails of eight test images.
• Figure 5: Visual quality comparisons between the proposed method and other image CS coding schemes in terms of R=0.2 and Bpp=0.2 on image Monarch.