Many optical systems are used for specific tasks such as classification. Of these systems, the majority are designed to maximize image quality for human observers. However, machine learning classification algorithms do not require the same data representation used by humans. We investigate the compressive optical systems optimized for a specific machine sensing task. Two compressive optical architectures are examined: an array of prisms and neutral density filters where each prism and neutral density filter pair realizes one datum from an optimized compressive sensing matrix, and another architecture using conventional optics to image the aperture onto the detector, a prism array to divide the aperture, and a pixelated attenuation mask in the intermediate image plane. We discuss the design, simulation, and trade-offs of these systems built for compressed classification of the Modified National Institute of Standards and Technology dataset. Both architectures achieve classification accuracies within 3% of the optimized sensing matrix for compression ranging from 98.85% to 99.87%. The performance of the systems with 98.85% compression were between an F / 2 and F / 4 imaging system in the presence of noise.
Images are often not the optimal data form to perform machine learning tasks such as scene classification. Compressive classification can reduce the size, weight, and power of a system by selecting the minimum information while maximizing classification accuracy.In this work we present designs and simulations of prism arrays which realize sensing matrices using a monolithic element. The sensing matrix is optimized using a neural network architecture to maximize classification accuracy of the MNIST dataset while considering the blurring caused by the size of each prism. Simulated optical hardware performance for a range of prism sizes are reported.
Alternative architectures for imaging devices which fuse the optical design with an algorithmic component enable inexpensive sensing systems optimized for specific classification tasks. Leveraging past work in task-specific compressive devices, this work seeks to improve upon previous designs of optical and algorithmic elements. We achieve this through use of genetic algorithms to enforce conditions upon the optimization phase of a computational imaging system. Through enforcement of binary sampling or discrete-valued outputs of a system measurement matrix, it is possible to simplify optical hardware design while achieving high task-specific performance.
Many optical systems are used for specific tasks such as classification. Of these systems, the majority are designed to maximize image quality for human observers; however, machine learning classification algorithms do not require the same data representation used by humans. In this work we investigate compressive optical systems optimized for a specific machine sensing task. Two compressive optical architectures are examined: An array of prisms and neutral density filters where each prism and neutral density filter pair realizes one datum from an optimized compressive sensing matrix, and another architecture using conventional optics to image the aperture onto the detector, a prism array to divide the aperture, and a pixelated attenuation mask in the intermediate image plane. We discuss the design, simulation, and tradeoffs of these compressive imaging systems built for compressed classification of the MNSIT data set. To evaluate the tradeoffs of the two architectures, we present radiometric and raytrace models for each system. Additionally, we investigate the impact of system aberrations on classification accuracy of the system. We compare the performance of these systems over a range of compression. Classification performance, radiometric throughput, and optical design manufacturability are discussed.
We investigate the feasibility of additively manufacturing optical components to accomplish task-specific classification in a computational imaging device. We report on the design, fabrication, and characterization of a non-traditional optical element that physically realizes an extremely compressed, optimized sensing matrix. The compression is achieved by designing an optical element that only samples the regions of object space most relevant to the classification algorithms, as determined by machine learning algorithms. The design process for the proposed optical element converts the optimal sensing matrix to a refractive surface composed of a minimized set of non-repeating, unique prisms. The optical elements are 3D printed using a Nanoscribe, which uses two-photon polymerization for high-precision printing. We describe the design of several computational imaging prototype elements. We characterize these components, including surface topography, surface roughness, and angle of prism facets of the as-fabricated elements.
Many optical systems are used for specific tasks such as classification. Of these systems, the majority are designed to maximize image quality for human observers; however, machine learning classification algorithms do not require the same data representation used by humans. In this work we investigate compressive optical systems optimized for a specific machine sensing task. Two compressive optical architectures are examined: An array of prisms and neutral density filters where each prism and neutral density filter pair realizes one datum from an optimized compressive sensing matrix, and another architecture using conventional optics to image the aperture onto the detector, a prism array to divide the aperture, and a pixelated attenuation mask in the intermediate image plane. We discuss the design, simulation, and tradeoffs of these compressive imaging systems built for compressed classification of the MNSIT data set. To evaluate the tradeoffs of the two architectures, we present radiometric and raytrace models for each system. Additionally, we investigate the impact of system aberrations on classification accuracy of the system. We compare the performance of these systems over a range of compression. Classification performance, radiometric throughput, and optical design manufacturability are discussed.
Optical remote sensing has become a valuable tool in many application spaces because it can be unobtrusive, search large areas efficiently, and is increasingly accessible through commercially available products and systems. In the application space of chemical, biological, radiological, nuclear, and explosives (CBRNE) sensing, optical remote sensing can be an especially valuable tool because it enables data to be collected from a safe standoff distance. Data products and results from remote sensing collections can be combined with results from other methods to offer an integrated understanding of the nature of activities in an area of interest and may be used to inform in-situ verification techniques. This work will overview several independent research efforts focused on developing and leveraging spectral and polarimetric sensing techniques for CBRNE applications, including system development efforts, field deployment campaigns, and data exploitation and analysis results. While this body of work has primarily focused on the application spaces of chemical and underground nuclear explosion detection and characterization, the developed tools and techniques may have applicability to the broader CBRNE domain.
In this effort, random noise data augmentation is compared to phenomenologically-inspired data augmentation for a target detection task, evaluated on the Digital Imaging and Remote Sensing Image Generation (DIRSIG) model "MegaScene" simulated hyperspectral dataset. Random data augmentation is commonly used in the machine learning literature to improve model generalization. While random perturbations of an input may work well in certain fields such as image classification, they can be unhelpful in other applications such as hyperspectral target detection. For instance, random noise augmentation may not be beneficial when the applied noise distribution does not match underlying physical signal processes or sensor noise. In the context of a low-noise sensor, augmentation mimicking material mixing and other practical spectral modulations is likely to be more effective when used to train a target detector. It is therefore important to utilize a data augmentation strategy that emulates the natural variability in observed spectra. To validate this claim, a small fully connected neural network architecture is trained using an ideal hemispheric reflectance materials dataset as a trivial baseline. That dataset is then augmented using Gaussian random noise and the model is retrained and again applied to MegaScene. Finally, augmentation is instead performed using phenomenological insight and used to retrain and reevaluate the model. In this work, the phenomenological augmentation implements only simple and commonly encountered spectral permutations, namely linear mixing and shadowing. Comparison is made between the augmented models and the baseline model in terms of low constant false alarm rate (CFAR) performance.
Advancements in machine learning (ML) and deep learning (DL) have enabled imaging systems to perform complex classification tasks, opening numerous problem domains to solutions driven by high quality imagers coupled with algorithmic elements. However, current ML and DL methods for target classification typically rely upon algorithms applied to data measured by traditional imagers. This design paradigm fails to enable the ML and DL algorithms to influence the sensing device itself, and treats the optimization of the sensor and algorithm as separate sequential elements. Additionally, this current paradigm narrowly investigates traditional images, and therefore traditional imaging hardware, as the primary means of data collection. We investigate alternative architectures for computational imaging systems optimized for specific classification tasks, such as digit classification. This involves a holistic approach to the design of the system from the imaging hardware to algorithms. Techniques to find optimal compressive representations of training data are discussed, and most-useful object-space information is evaluated. Methods to translate task-specific compressed data representations into non-traditional computational imaging hardware are described, followed by simulations of such imaging devices coupled with algorithmic classification using ML and DL techniques. Our approach allows for inexpensive, efficient sensing systems. Reduced storage and bandwidth are achievable as well since data representations are compressed measurements which is especially important for high data volume systems.
Channeled spectropolarimetry measures the spectrally resolved Stokes parameters. A key aspect of this technique is to accurately reconstruct the Stokes parameters from a modulated measurement of the channeled spectropolarimeter. The state-of-the-art reconstruction algorithm uses the Fourier transform to extract the Stokes parameters from channels in the Fourier domain. While this approach is straightforward, it can be sensitive to noise and channel cross-talk, and it imposes bandwidth limitations that cut o high frequency details. To overcome these drawbacks, we present a reconstruction method called compressed channeled spectropolarimetry. In our proposed framework, reconstruction in channeled spectropolarimetry is an underdetermined problem, where we take N measurements and solve for 3N unknown Stokes parameters. We formulate an optimization problem by creating a mathematical model of the channeled spectropolarimeter with inspiration from compressed sensing. We show that our approach o ers greater noise robustness and reconstruction accuracy compared with the Fourier transform technique in simulations and experimental measurements. By demonstrating more accurate reconstructions, we push performance to the native resolution of the sensor, allowing more information to be recovered from a single measurement of a channeled spectropolarimeter.