Optimization of the architecture of road defect detection models based on artificial neural networks to increase prediction accuracy while reducing computing power requirements

MINIATURA 8

Funding Organization: National Science Centre

Project title: Optimization of the architecture of road defect detection models based on artificial neural networks to increase prediction accuracy while reducing computing power requirements

Agreement number: DEC-2024/08/X/ST6/00610

Project implementation period: 06.09.2024 - 05.09.2025

Principal Investigator: dr inż. Paweł Rafał Tomiło

Project value: 11 055,00 PLN

Funds granted for Lublin University of Technology: 11 055,00 PLN

Abstract: The research is aimed at optimizing or creating an artificial neural network model architecture for object detection. Over the past few years, the development of artificial intelligence algorithms has allowed significant advances in object detection, which is used in many areas of life, from industry to medicine. Through the use of advanced machine learning techniques, such as deep learning, it has become possible to create neural network models capable of accurately identifying and locating objects in images or video streams. As a result, we can now see increasing efficiency in automatic recognition of faces, vehicles or other objects, with applications in surveillance systems, public safety or autonomous cars. In addition, the development of neural network architectures enables more efficient processing of data and reduction of computational costs, making artificial intelligence-based technologies increasingly accessible and widespread in various areas of life.The main objective of the research is to develop an optimized neural network architecture with higher accuracy and lower computational power requirements through the use of layers such as LSK, Involution or changing the backbone architecture and neck in detection models, among others. Developing an optimized neural network architecture may require a lot of experimentation and analysis of various factors affecting model performance, such as the type of training data, network size, method of regularization or optimization techniques. In addition, understanding the mechanisms of the different layers of the network and adjusting their parameters accordingly are key to achieving the desired results. Through systematic research and an iterative process of improvement, in the research I will strive to create an object detection model that is not only effective, but also efficient under conditions of limited computing resources - a single-board minicomputer. Particular emphasis will be placed on the efficient use of computing resources and the integration of the model with technical systems - data processing equipment. In this regard, it will also be important to adapt the model architecture to practical requirements, such as processing time or accuracy in the context of analysis of images obtained from developed system.The developed artificial neural network model will find application in the detection of road surface defects, which will directly contribute to improving the quality of road infrastructure management and optimize diagnostic and repair processes. In addition, its implementation will improve the road asset management process, enabling more precise planning and implementation of maintenance activities. By automatically monitoring the condition of the pavement at the vehicle level, road managers will be able to respond more quickly to emerging defects, which will not only minimize repair costs, but also extend the life of the pavement, increasing safety and comfort for users.