TL;DR
This paper introduces a novel dual-optical path imaging system for compressive sensing (CS) that simultaneously captures CS measurements and ground-truth images. By leveraging this dual-path approach and a deep learning framework with separate filter and reconstruction networks, the system effectively mitigates hardware-induced noise, leading to significantly improved image reconstruction quality compared to existing methods.
What Problem Does This Paper Address?
Traditional compressive sensing (CS) imaging systems face a critical challenge: the quality of reconstructed images is severely hampered by noise originating from the optical path and electronic circuits. This noise is amplified during the high-gain signal processing required for CS, leading to blurry and artifact-ridden images. Furthermore, acquiring clean, labeled data for training deep learning models in CS is difficult because the sampling process inherently compresses and potentially distorts the signal. The paper aims to address these limitations by proposing a system that not only acquires CS measurements but also provides simultaneously captured, high-quality ground-truth images for robust training, and employs a deep learning approach to actively filter out hardware noise.
Methods
The core innovation lies in a dual-optical path imaging system. This system utilizes the bidirectional reflection capabilities of Digital Micromirror Devices (DMDs), which can reflect light at ±12 degrees. One path is dedicated to the standard CS sampling process, where the DMD modulates the light according to a measurement matrix to generate compressed measurements. The second, parallel optical path captures the scene's light directly onto a CMOS sensor, effectively providing a high-quality, noise-free ground-truth image of the same scene from the same viewing angle. This dual-path setup allows for the simultaneous acquisition of both the noisy CS measurements and the clean reference images, crucial for training.
The acquired data is then fed into a deep learning framework consisting of two main components:
- Filter Network: This network, built using convolutional neural networks (CNNs), is trained to specifically denoise the compressed CS measurements. It learns to identify and remove the hardware-induced noise without significantly altering the underlying signal information.
- Reconstruction Network: Following the filter network, a reconstruction network (composed of fully connected layers) takes the denoised measurements and reconstructs the final image.
The training strategy involves three parts: pre-training the filter network, pre-training the reconstruction network, and then joint training of both networks with the pre-trained weights, using data sampled from the dual-optical path system.
Datasets
The paper does not explicitly mention using publicly available datasets for training or evaluation. Instead, it introduces a novel data acquisition method using its proposed dual-optical path system. The system captures a total of 44 images from the device, with 8 of these reserved for testing. For training the reconstruction network, approximately 14,000 simulated training data sets were used, likely derived from existing image datasets. For joint training, 36 datasets sampled directly from their device were utilized.
Key Insights and Ideas
- Dual-Path for Ground Truth: The most significant insight is the realization that the ±12 degree reflection capability of DMDs can be exploited to create a parallel optical path for capturing ground-truth images. This elegantly solves the problem of obtaining labeled data for CS reconstruction without relying on external datasets or complex simulation assumptions.
- Hardware Noise as a Learnable Target: Instead of solely focusing on algorithmic reconstruction, the paper treats hardware noise as a primary target for a dedicated deep learning filter. This adaptive filtering approach directly tackles the root cause of poor image quality in practical CS systems.
- Synergy of Filtering and Reconstruction: The paper demonstrates that pre-filtering the noisy measurements before reconstruction significantly boosts the performance of not only their proposed reconstruction network but also traditional CS reconstruction algorithms. This highlights the general applicability of their noise-filtering approach.
- High Sampling Rate Challenges: The results suggest that while higher sampling rates (measurement rates) are generally desirable, they can exacerbate cumulative errors due to longer acquisition times, potentially leading to worse reconstruction quality for some traditional methods. Their proposed method appears more robust to this.
Key Results
The proposed dual-optical path CS imaging system, coupled with the filter and reconstruction networks, demonstrates significant improvements in image reconstruction quality.
- Superior Performance: The method consistently achieves the best PSNR and SSIM scores across all tested measurement rates (MRs) compared to established algorithms like TVAL3, NLR-CS, and D-AMP. For instance, at a 25% MR, their method achieved PSNR values significantly higher than others, with specific examples showing improvements of over 10 dB.
- Effective Noise Suppression: The filter network demonstrably reduces low-frequency noise in the sampled signals, making them much closer to the original signal. This is quantified by a significant reduction in the mean square error between the sampled and filtered signals.
- Improved Reconstruction for Others: Applying their filter network to the noisy measurements before feeding them into traditional algorithms like TVAL3, NLR-CS, and D-AMP resulted in substantial improvements in PSNR and SSIM, with some algorithms seeing an increase of up to 7 dB.
- Speed: The neural network-based reconstruction, accelerated by GPU, achieves an average reconstruction time of approximately 0.02 seconds per image, a significant speedup compared to traditional CPU-based methods.
Limitations and Future Work
- Data Volume: The paper acknowledges that the proposed system generates a large amount of image data during operation, which could be a practical consideration.
- Future Work: The authors suggest that future research should focus on developing techniques to continuously update the filter network online, leveraging the newly generated image data. This would allow the system to adapt to changing environmental conditions or hardware drift over time.
Summary updated