Pattern recognition in images based on artificial neural networks and complex systems: from handcrafted descriptor extraction to automated learning

Abstract

Pattern recognition in images is a key topic of computer vision, involving the analysis of visual features (such as texture, color, and shape) and having various applications. Currently, there are two main paradigms of methods, each with advantages and disadvantages: the classic, which uses manual descriptors and requires greater domain knowledge, and approaches based on machine learning with Artificial Neural Networks (ANNs). Classic methods use mathematical or statistical models and are useful in scenarios with limited data, hardware, and a greater need for interpretability. On the other hand, learning-based approaches, such as convolutional neural networks, are gaining popularity due to their ability to learn relevant features from large volumes of data and high classification performance. This project proposes the study and development of methods based on both paradigms, using complex systems techniques and ANNs to develop methods that improve pattern recognition tasks.More specifically, the focus will be on two complex systems approaches: complex networks and deterministic walks. These approaches accurately describe the irregularity or homogeneity of structures in images, relevant for natural and artificial vision, assisting in the analysis of images from nature or stemming from non-linear phenomena.In ANNs, Randomized Neural Networks, Convolutional Neural Networks, and Vision Transformers will be studied, exploring the potential of each architecture in pattern recognition in images and complex systems. Thus, this project can also explore the symbiosis between ANNs and complex systems, investigating and interpreting neural networks from the perspective of complex systems, making the use of these approaches for pattern recognition a natural line of research, as highlighted by some studies. In addition to the theoretical front, as a way to contribute to other areas, another goal is to apply the developed methods to real problem data generated by partners, especially in images of materials (sensors and biosensors), plants, and medical fields. Therefore, it is emphasized that the project seeks to create an interaction between application and theory, testing generic methods in specific problems and, at the same time, allowing these problems to inspire new theoretical lines, advancing knowledge on both fronts. (AU)