TL;DR
This paper introduces a novel "Dual-Channel Reconstruction Network" (DC-Net) module for image compressive sensing (CS). By combining residual and dense blocks, their proposed RDC-Net and FDC-Net achieve superior reconstruction quality and faster processing times compared to existing deep learning and traditional CS methods, particularly for images acquired with a fully connected measurement matrix.
What Problem Does This Paper Address?
Compressive sensing (CS) allows for acquiring signals (like images) with fewer measurements than traditional methods, but reconstructing the original signal from these sparse measurements is computationally intensive and often results in suboptimal quality. Existing deep learning-based CS reconstruction methods, while faster than traditional iterative optimization techniques, still face challenges in achieving both high reconstruction accuracy and computational efficiency. Specifically, the way measurements are acquired (the measurement matrix) significantly impacts reconstruction quality, and current methods often struggle to effectively leverage the information captured by different types of measurement matrices.
Methods
The core innovation is the Dual-Channel Reconstruction Network (DC-Net) module. This module is designed to improve upon a preliminary reconstructed image obtained from a simple fully connected layer. The DC-Net module itself comprises two parallel channels:
- Residual Block Channel: Utilizes two residual blocks, known for their ability to capture rich image features and improve reconstruction quality by mitigating the vanishing gradient problem through identity shortcut connections.
- Dense Block Channel: Employs one dense block, which excels at feature reuse and parameter efficiency, thereby expediting computation.
These two channels work in tandem, with the residual channel focusing on high-level features and the dense block channel on low-level features. This dual-channel approach aims to balance reconstruction quality and computational speed.
The paper proposes two main networks built using this DC-Net module:
- RDC-Net: Uses a traditional random Gaussian measurement matrix for acquiring CS measurements.
- FDC-Net: Utilizes a fully connected measurement matrix, which the authors suggest captures original image information more effectively.
Both networks first use a fully connected layer to generate a preliminary reconstruction, which is then refined by the DC-Net module.
Datasets
The paper uses publicly available datasets for training and testing:
- Training Data: Luminance components of images from a dataset (referenced as [16]) were used. 33x33 image patches with a stride of 14 were extracted.
- Test Data: Similar to training, 33x33 patches with a stride of 14 were extracted from images in the same dataset [16].
- Scalability Test: The large-scale ImageNet validation dataset [52] was used to evaluate the generalization ability of the proposed networks.
Key Insights and Ideas
- Synergy of Residual and Dense Blocks: The paper demonstrates that combining residual blocks (for feature richness) and dense blocks (for computational efficiency) in a dual-channel architecture leads to superior performance compared to using either type of block alone or deeper single-channel networks.
- Effectiveness of Fully Connected Measurement Matrix: FDC-Net, which uses a learned fully connected measurement matrix, consistently outperforms RDC-Net and other state-of-the-art methods, especially at higher undersampling rates (MR=0.25), suggesting that this measurement strategy can capture more informative data for reconstruction.
- Shallow Network Advantage: The proposed dual-channel networks are relatively shallow compared to some existing deep learning methods, yet they achieve better results, highlighting the efficiency of their architectural design.
- Robustness to Noise: The FDC-Net shows a slower performance degradation under Gaussian noise compared to DR2-Net, indicating better robustness.
Key Results
- PSNR Improvement: FDC-Net achieved the highest mean PSNR values across various measurement rates (MRs), notably showing a 2.3 dB improvement over the second-highest value at MR = 0.25.
- Visual Quality: Figures 3-5 visually demonstrate that the proposed networks, particularly FDC-Net, produce clearer textures, sharper edges, and fewer artifacts compared to DR2-Net.
- Reduced Runtime: Table 5 shows that FDC-Net has a slightly lower runtime (e.g., 0.0566s for MR=0.04 on a 512x512 image) compared to DR2-Net (0.0676s), making it more suitable for real-time applications.
- Generalization: On the ImageNet validation dataset, RDC-Net and FDC-Net showed significant PSNR improvements over DR2-Net, indicating good generalization capabilities.
- Convergence: Figure 7 illustrates that FDC-Net converges faster with smaller training and test errors than DR2-Net.
Limitations and Future Work
- Block-wise Processing Artifacts: While BM3D is used to mitigate artifacts from block-wise processing, it's acknowledged as a post-processing step. Future work could explore end-to-end methods that inherently avoid such artifacts.
- Dataset Specificity: The experiments were conducted on specific datasets. Further validation on a wider variety of image types and real-world CS scenarios would be beneficial.
- Measurement Matrix Learning: While FDC-Net uses a learned fully connected matrix, exploring more sophisticated learned measurement matrices could potentially yield even better results.
- Computational Cost of FDC-Net: Although FDC-Net is faster than DR2-Net, the computational cost of learning the fully connected measurement matrix itself might be a consideration for extremely large-scale applications.
Summary updated