Computationally efficient signal processing of image sequences for thermal inspection of additive manufacturing

This thesis presents a unified, physics-informed framework for the spatio-temporal super-resolution of thermal imaging, motivated by the limitations of commercial cameras in monitoring dynamic additive manufacturing (AM) processes. Conventional infrared systems don't simultaneously achieve high spatial and temporal resolution, constraining their ability to capture rapid thermal transients that govern part quality.

The challenge is addressed by reformulating thermal image enhancement as a state-estimation problem grounded in the heat diffusion equation. A discretised state-space model enables the use of Kalman filtering and Rauch–Tung–Striebel (RTS) smoothing to reconstruct high-resolution, high-frame-rate thermal fields from spatially and temporally downsampled observations. This physics-informed approach produces reconstructions that preserve sharp spatial details and transient dynamics, significantly outperforming interpolation-based methods in both accuracy and stability.

To overcome the computational burden of standard Kalman filtering, a distributed estimation architecture was developed through the Reduced-Update Partition-Based Kalman Filter (RU-PBKF) and Reduced-Update Partition-Based Smoother (RU-PBS). By partitioning the image into locally interacting subsystems, the framework exploits the inherent locality of heat diffusion to achieve scalability, reducing computational complexity while maintaining estimation fidelity.

The framework was further extended to accommodate rolling shutter imaging, using a time-varying observation model and state-augmentation strategy to correct temporal misalignment and perform temporal super-resolution. Finally, a Generalised Likelihood Ratio Test (GLRT) with a CUSUM-type accumulator was integrated for online detection, estimation, and compensation of unknown thermal inputs, enabling adaptive response to unmodelled heat events.

Comprehensive simulations demonstrate that the proposed framework achieves accurate, scalable, and robust reconstruction of transient thermal phenomena. The framework provides a foundation for real-time, physics-consistent monitoring in additive manufacturing and other thermally driven processes.