- author = "Illya Olegovich Bakurov",
- title = "Soft computing for Ill Posed Problems in Computer Vision",
- school = "NOVA Information Management School (NOVA IMS), NOVA University, Lisbon",
- year = "2022",
- address = "Portugal",
- month = "19 " # sep,
- keywords = "genetic algorithms, genetic programming, ANN, Evolutionary Computation, Swarm Intelligence, Ensemble Learning, Stacking, Computer Vision, Full Reference Image Quality Assessment, Semantic Segmentation",
- language = "English",
-
URL = "
http://hdl.handle.net/10362/144500",
-
URL = "
https://run.unl.pt/bitstream/10362/144500/1/D0071.pdf",
- size = "177 pages",
- abstract = "Soft computing (SC) includes computational techniques that are tolerant of approximations, missing information, and uncertainty, and aim at providing effective and efficient solutions to problems which may be unsolvable, or too time-consuming to solve, with exhaustive techniques. SC has found many applications in various domains of research and industry, including computer vision (CV). This dissertation focuses on tasks of full reference image quality assessment (FR-IQA) and fast scene understanding (FSU). The former consists of assessing images visual quality in regard to some pristine reference. The latter consists of classifying each pixel of a scene assuming a rapidly changing environment like, for instance, in a self-driving car. The current state-of-the-art (SOTA) in both FR-IQA and FSU rely upon convolutional neural networks (CNNs), which can be seen as a computational metaphor of the human visual cortex. Although CNNs achieved unprecedented results in many CV tasks, they also present several drawbacks: massive amounts of data and processing resources for training; the difficulty of outputs interpretation; reduced usability for compact battery-powered devices... This dissertation addresses FR-IQA and FSU using SC techniques other than CNNs. Initially, we created a flexible and efficient library to support our endeavors; it is publicly available and implements a wide range of metaheuristics to solve different problems. Then, we used swarm and evolutionary computation to optimize the parameters of several traditional FR-IQA measures (FR-IQAMs) that integrate the so called structural similarity paradigm; the novel parameters improve measures’ precision without affecting their complexity. Afterward, we applied genetic programming (GP) to automatically formulate novel FR-IQAMs that are simultaneously simple, accurate, and interpretable. Lastly, we used GP as a meta-model for stacking efficient CNNs for FSU; the approach allowed us to obtain simple and interpretable models that did not exceed processing preconditions for real-time applications while achieving high levels of precision.",
-
notes = "In English
Supervisors: Mauro Castelli, Leonardo Vanneschi",
