Deep learning-based systems have shown superior performance in many artificial intelligence tasks in the last decade. Imaging systems are particularly well suited to benefit from deep learning with the use of Convolutional Neural Networks (CNNs). CNNs and other deep learning methods are now the preferred approach for nearly all computer vision and pattern recognition challenges, including image classification, object detection, and biometric recognition. Not only are CNNs used with images captured in the visible spectrum wavelengths, but they have also been used with varying success on images captured in other wavelengths, most often in the infrared (IR) spectrum. The different regions of the IR band each have their own unique advantages, providing researchers with new ways to investigate computer vision problems in challenging scenarios that may be impossible to solve in the visible spectrum. Although deep learning-based imaging systems have been used successfully for a variety of tasks in both the visible and IR spectrums, many challenges remain, especially for biometric recognition applications. Studies show that a performance gap exists between the image-based systems that operate in different spectrums. These challenges can generally be attributed to the large appearance variations of images captured in the different spectrums and a limited availability of high quality data in the non-visible bands.

In this dissertation I address several of the gaps that exist in the open literature for imaging systems that use multiple spectrums, particularly as they relate to biometric systems and object detection. The dissertation includes an introduction and background information for each of the challenges covered in this work and a systematic review of the relevant literature. Methods are proposed to address facial attribute analysis in the visible and middle-wave IR bands, facial landmark detection and recognition via image synthesis using the visible and passive (middle and long wave IR) bands, and detection of firearms from surveillance videos in the visible band. The proposed approaches and results from this dissertation provide practical solutions and analysis for a variety of imaging system challenges and provide helpful insights and directions for future research.