From sensors to machine vision systems: Exploring machine vision, computer vision and machine learning with hyperspectral imaging applications

Abstract

The hypothesis of this study is “The machine vision systems should be designed, built and evaluated through the machine vision fundamental phases.” The dissertation defines the fundamentals inspired by the literature and shows how to use them and design a machine vision system from the sensor level to the analysis. The study covers the hardware and software designing phases, continuing to the data pre-processing, transformation and analysis phases. The conducted research consists of six published articles. Three have a data-analytical point of view concentrating on developing and testing new versions of a distance-based machine learning method Minimal Learning Machine (MLM). In contrast, the rest of the articles introduces a novel concept of 3D hyperspectral imaging and convolutional neural networks for detecting, classifying and delineating skin cancer on complex skin surfaces. The skin cancer imager concept includes devices, user interfaces, pre-processing, transformation and analysis. As the main results, we introduce the machine vision fundamentals with three effective variations from the MLM, which is suitable for hyperspectral imaging anomaly detection and classification tasks in real-time applications. Other outcomes were a novel hyperspectral imaging concept that can reach complex skin surfaces, opening the road for future optical biopsy. As skin cancers are the world’s third most common type of cancer, an optical biopsy can reduce diagnosis and treatment costs and save lives through early, accurate detection. The results confirm that by developing machine vision systems according to the application and paying attention to the machine vision fundamentals, it is possible, for example, to influence the system’s computational complexity and improve the system’s results. By understanding the multi-directional relations between the fundamental phases of a machine vision system, we can affect the overall performance. For instance, by designing a machine vision system that collects only the necessary data and uses optimised fast computational methods, we can affect the system’s efficiency, energy need and operating costs.