Image super-resolution for consumer thermal cameras

This thesis attempts to solve a problem currently faced by consumer-grade thermal cameras. Consumers have grown accustomed to the high pixel counts and high resolution offered by the modern smartphone, while the performance of thermal cameras have not yet reached the same point. This problem is ideal for super-resolution, though most existing approaches to super-resolution are not suitable for application in consumer devices. In this work a method of super-resolution was developed using multiple input images which do not fulfil the requirements of the sampling theorem. This method uses sub-pixel shifts between frames to recover high spatial frequency information from below the diffraction-limit that would otherwise appear as aliasing artefacts in each image. The technique developed operates by convolution of offset Sinc kernels that are adjusted by the shift of each frame with internally padded versions of input images. This is shown for images that undergo global shifts alone before support for more complicated scene or object motion is included. A locally variable convolution step was developed to allow for super-resolution using convolution kernels that vary based on knowledge of local image shifts to the level of each pixel. Consideration was made for computation time, robustness against errors in shift measurement and the reduction in required computation using this technique. Additional original contributions made in this work include the design and construction of thermal imaging targets that allow for assessment of thermal camera performance. While many standard imaging targets exist to test visible camera resolution, these do not transfer to testing of thermal cameras, which rely on the emissivity of objects for imaging. Equivalent resolution targets were designed and fabricated using 3D printing for this work. A cooled high-contrast background was also developed, along with a heated wire target to allow for measurement of the point spread function of thermal camera systems. Finally, a method was developed for image registration under conditions of additive or multiplicative fixed-pattern noise. While it is possible to compensate for fixed-pattern noise in some imaging conditions, this is not always possible, requiring a method to compensate for fixed patterns during registration.

Funder

Irish Research Council

Publisher

NUI Galway

Rights

Attribution-NonCommercial-NoDerivs 3.0 Ireland

cbn

Collections

University of Galway Theses (PhD Theses)