Efficient Models and Algorithms for Image Processing for Industrial Applications

Image processing and computer vision are now part of our daily life and allow artificial intelligence systems to see and perceive the world with a visual system similar to the human one. In the quest to improve performance, computer vision algorithms reach remarkable computational complexities. The high computational complexity is mitigated by the availability of hardware capable of supporting these computational demands. However, high-performance hardware cannot always be relied upon when one wants to make the research product usable. In this work, we have focused on the development of computer vision algorithms and methods with low computational complexity but high performance. The first approach is to study the relationship between Fourier-based metrics and Wasserstein distances to propose alternative metrics to the latter, considerably reducing the time required to obtain comparable results. In the second case, instead, we start from an industrial problem and develop a deep learning model for change detection, obtaining state-of-the-art performance but reducing the computational complexity required by at least a third compared to the existing literature.