Download Application of Machine Learning to Optical Sensing and Imaging Book in PDF, ePub and Kindle
Recent years have witnessed significant advancements in optical sensing and imaging techniques. To effectively interpret complex data acquired through these techniques and accurately extract information from detectors, machine learning has emerged as a promising solution. Machine learning enables automatic learning of the relationship between raw data and desired outputs, without the need for complete and explicit physics-based models. This data-driven approach presents opportunities for making inferences on material properties, solving inverse problems in the area of optical sensing and imaging. However, the current majority of machine learning methods applied in optical systems primarily serve as post-processing tools to enhance automation and improve the signal-to-noise ratio after data acquisition with standard optical systems. This approach often utilizes precision optics that can be bulky and expensive, along with typical machine learning algorithms that may not fully exploit the underlying physics. The separation of optics and algorithms in design and optimization limits the potential for integration and performance improvement. Consequently, an area of fruitful research lies in deeply integrating machine learning into the design of efficient optical hardware systems, optimizing and streamlining their performance. This dissertation aims to demonstrate how machine learning can contribute to the design of cost-effective and portable optical devices by leveraging minimal optical components in conjunction with powerful learning models. The proposed approach adopts a holistic perspective in designing optical sensing systems. By relieving the burden on optics, simpler and more affordable optical components can be utilized. Moreover, optical domain knowledge can be effectively employed to custom design efficient machine learning algorithms. The dissertation is divided into three main parts, exploring different spaces: spectral, polarization, spatial-temporal, and more. Each part focuses on enhancing and improving optical system design through the application of machine learning techniques. The first part investigates label-free bio-imaging, specifically addressing the urgent need for rapid bacterial diagnostics. Traditional gold-standard methods for bacterial diagnostics are often time-consuming, leading to delays in prescribing appropriate treatments for antimicrobial resistance. To accelerate antibacterial susceptibility testing (AST), dynamic laser speckle imaging methods are introduced in Chapter 2. Speckle images are captured, and a machine learning model is employed to track and analyze the dynamic patterns of bacteria, predicting the minimum inhibitory concentration (MIC) within a significantly reduced time frame of one hour. Furthermore, Chapter 3 proposes matched and Principal Component Analysis (PCA) Raman as a potential means to reduce the time required for bacterial identification. It demonstrates the enhancement of Raman scattering through the modulation of the excitation laser and the customization of spectral filters. Machine learning methods guide the hardware-level design of these filters to optimize efficiency and selectivity. Raman sensing is showed for classifying bacteria samples, potentially aiding in rapid bacterial identification in solution. Raman imaging is also demonstrated through the scanning of polystyrene spheres and yeast samples. The second part of the dissertation explores the application of deep learning as a data-driven method to solve inverse problems and enable real-time imaging. Chapter 4 presents a deep learning-based non-line-of-sight (NLOS) imaging system developed to reconstruct occluded objects from scattering surfaces. The system can be trained using only handwritten digits, yet it exhibits the capability to reconstruct patterns beyond the training set, including physical objects and real-time cartoon videos. By utilizing an ordinary camera and incoherent light source, this approach enables a cost-effective and real-time NLOS imaging system without the need for an explicit physical model of light transport. Deep learning is further applied to multidimensional imaging in Chapter 5, where intensity, polarization, and spectrum can be measured. In non-line-of-sight scenarios, where direct access to the object is unavailable, the scattering surface provides the only scrambled information. Conversely, in hyperspectral polarimetric imaging, direct access to the object allows for the artificial design of metasurfaces to encode light field information. By capturing images with a normal camera, encoded by a metasurface sensitive to spectral and polarization properties, the deep learning model accurately reconstructs spectral and polarization parameters for both laser and white light illumination. Demonstrations include the reconstruction of real objects, with video-rate reconstruction at over 10 frames per second. The third part of the dissertation takes a step further and explores the role of machine learning in ultrafast imaging at trillions frame per second. Chapter 6 introduces collinear frequency-resolved optical gating (FROG) for ultrafast optical pulse reconstruction. Machine learning methods are leveraged to achieve pulse retrieval at the femtosecond level, performing over 200 reconstructions per second, compared to the minutes required by traditional iterative algorithms. This advancement potentially enables high-speed imaging in nanostructures. The machine learning model is trained using simulation data that incorporates noise to replicate experimental conditions. Results demonstrate the model's ability to accurately reconstruct both simulated and experimental pulses with high precision. Additionally, Chapter 7 presents a theoretical study on ultrafast imaging with spatiotemporal mask and deep learning. Ultrafast events are encoded by spatiotemporal mask and detected using a normal camera. A deep learning model is designed to reconstruct ultrafast event sequences that are as close as 200 fs apart, utilizing single-shot images. In conclusion, this dissertation highlights the potential of machine learning to optimize optical hardware systems for various applications, including label-free bio-imaging, real-time imaging, and ultrafast imaging. By integrating machine learning techniques into optical design, cost-effective and portable devices can be developed, ushering in advancements in optical sensing and imaging technologies.