skip to main content

Title: End-to-End Optimized Adversarial Deep Compressed Super-Resolution Imaging via Pattern Scanning
We propose an end-to-end optimized adversarial deep compressed imaging modality. This method exploits the adversarial duality of the sensing basis and sparse representation basis in compressed sensing framework and shows solid super-resolution results.
; ;
Award ID(s):
Publication Date:
Journal Name:
OSA Imaging and Applied Optics Congress 2021 (3D, COSI, DH, ISA, pcAOP)
Page Range or eLocation-ID:
Sponsoring Org:
National Science Foundation
More Like this
  1. Multi-channel data acquisition of bio-signals is a promising technology that is being used in many fields these days. Compressed sensing (CS) is an innovative approach of signal processing that facilitates sub-Nyquist processing of bio-signals, such as an electrocardiogram (ECG) and electroencephalogram (EEG). This strategy can be used to lower the data rate to realize ultra-low-power performance, As the count of recording channels increase, data volume is increased resulting in impermissible transmitting power. This paper presents the implementation of a CMOS-based front-end design with the CS in the standard 180 nm CMOS process. A novel pseudo-random sequence generator is proposed, whichmore »consists of two different types of D flip-flops that are used for obtaining a completely random sequence. The power consumed by the bio-signal amplifier block is 2.35 μW. The SAR-ADC block that is designed to digitize the amplified signal consumes 277 μW of power and the power consumption value of the pseudo-random bit sequence generator (PRBS) is 344.2μW. The sampling rate of PRBS block is 611.76 Kbps. We have considered collecting neural data from the 32 channels, and achieved an 8.5X compression rate. The low power consumption per channel confirms the importance of the proposed approach for multiple channel high-density neural interfaces.« less
  2. Glow discharge optical emission spectroscopy elemental mapping (GDOES EM), enabled by spectral imaging strategies, is an advantageous technique for direct multi-elemental analysis of solid samples in rapid timeframes. Here, a single-pixel, or point scan, spectral imaging system based on compressed sensing image sampling, is developed and optimized in terms of matrix density, compression factor, sparsifying basis, and reconstruction algorithm for coupling with GDOES EM. It is shown that a 512 matrix density at a compression factor of 30% provides the highest spatial fidelity in terms of the peak signal-to-noise ratio (PSNR) and complex wavelet structural similarity index measure (cw-SSIM) whilemore »maintaining fast measurement times. The background equivalent concentration (BEC) of Cu I at 510.5 nm is improved when implementing the discrete wavelet transform (DWT) sparsifying basis and Two-step Iterative Shrinking/Thresholding Algorithm for Linear Inverse Problems (TwIST) reconstruction algorithm. Utilizing these optimum conditions, a GDOES EM of a flexible, etched-copper circuit board was then successfully demonstrated with the compressed sensing single-pixel spectral imaging system (CSSPIS). The newly developed CSSPIS allows taking advantage of the significant cost-efficiency of point-scanning approaches (>10× vs. intensified array detector systems), while overcoming (up to several orders of magnitude) their inherent and substantial throughput limitations. Ultimately, it has the potential to be implemented on readily available commercial GDOES instruments by adapting the collection optics.« less
  3. Recent advances in machine learning, especially techniques such as deep neural networks, are enabling a range of emerging applications. One such example is autonomous driving, which often relies on deep learning for perception. However, deep learning-based perception has been shown to be vulnerable to a host of subtle adversarial manipulations of images. Nevertheless, the vast majority of such demonstrations focus on perception that is disembodied from end-to-end control. We present novel end-to-end attacks on autonomous driving in simulation, using simple physically realizable attacks: the painting of black lines on the road. These attacks target deep neural network models for endto-endmore »autonomous driving control. A systematic investigation shows that such attacks are easy to engineer, and we describe scenarios (e.g., right turns) in which they are highly effective. We define several objective functions that quantify the success of an attack and develop techniques based on Bayesian Optimization to efficiently traverse the search space of higher dimensional attacks. Additionally, we define a novel class of hijacking attacks, where painted lines on the road cause the driverless car to follow a target path. Through the use of network deconvolution, we provide insights into the successful attacks, which appear to work by mimicking activations of entirely different scenarios. Our code is available on« less
  4. End-to-end data-driven image compressive sensing reconstruction (EDCSR) frameworks achieve state-of-the-art reconstruction performance in terms of reconstruction speed and accuracy. However, due to their end-to-end nature, existing EDCSR frameworks can not adapt to a variable compression ratio (CR). For applications that desire a variable CR, existing EDCSR frameworks must be trained from scratch at each CR, which is computationally costly and time-consuming. This paper presents a generic compression ratio adapter (CRA) framework that addresses the variable compression ratio (CR) problem for existing EDCSR frameworks with no modification to given reconstruction models nor enormous rounds of training needed. CRA exploits an initialmore »reconstruction network to generate an initial estimate of reconstruction results based on a small portion of the acquired measurements. Subsequently, CRA approximates full measurements for the main reconstruction network by complementing the sensed measurements with resensed initial estimate. Our experiments based on two public image datasets (CIFAR10 and Set5) show that CRA provides an average of 13.02 dB and 5.38 dB PSNR improvement across the CRs from 5 to 30 over a naive zero-padding approach and the AdaptiveNN approach(a prior work), respectively. CRA addresses the fixed-CR limitation of existing EDCSR frameworks and makes them suitable for resource-constrained compressive sensing applications.« less
  5. Compressed sensing (CS) as a new data acquisition technique has been applied to monitor manufacturing processes. With a few measurements, sparse coefficient vectors can be recovered by solving an inverse problem and original signals can be reconstructed. Dictionary learning methods have been developed and applied in combination with CS to improve the sparsity level of the recovered coefficient vectors. In this work, a physics-constrained dictionary learning approach is proposed to solve both of reconstruction and classification problems by optimizing measurement, basis, and classification matrices simultaneously with the considerations of the application-specific restrictions. It is applied in image acquisitions in selectivemore »laser melting (SLM). The proposed approach includes the optimization in two steps. In the first stage, with the basis matrix fixed, the measurement matrix is optimized by determining the pixel locations for sampling in each image. The optimized measurement matrix only includes one non-zero entry in each row. The optimization of pixel locations is solved based on a constrained FrameSense algorithm. In the second stage, with the measurement matrix fixed, the basis and classification matrices are optimized based on the K-SVD algorithm. With the optimized basis matrix, the coefficient vector can be recovered with CS. The original signal can be reconstructed by the linear combination of the basis matrix and the recovered coefficient vector. The original signal can also be classified to identify different machine states by the linear combination of the classification matrix and the coefficient vector.« less