Compressive imaging systems and algorithms to extend machine vision beyond the visible spectrum
Kelly, Kevin F
Doctor of Philosophy
Machine vision finds its importance in today’s most revolutionary technologies from artificial intelligence that surpasses humans in playing Go and chess to automobiles that drive themselves. For many of these tasks the key component that makes superior machine vision possible is the image sensor technology development that has paralleled the equally rapid development of processing power. However there is still a dilemma between the pursuit of higher resolution images that require a focal plane array (FPA) with more pixels on the front end, and the demands on acquisition for embedded systems restrained by power, transmission bandwidth, and storage. To overcome these challenges, the works presented in this thesis aim to seek solutions in solving particular machine vision tasks with compressive imaging system and advanced algorithms. The first strategy focused on achieving more robust infrared object classification utilizing measurements directly from the single-pixel camera without reconstruction with a multiscale compressive matched filter algorithm. Secondly, a multi-pixel hybrid optical convolutional neural network machine vision system was designed and validated to perform high-speed infrared object detection. Lastly, an approach to accomplish super-resolution beyond the resolutions of both the spatial light modulator and FPA in a compressive imaging system will be demonstrated by exploiting a coded point spread function to obtain sub-pixel information. Both simulation and experiment results were presented and analyzed to demonstrate the result of super-resolving an image with 4 times more of it original resolution. Resolving images beyond 4 times of their original resolutions is also possible by extending the idea of this work.