Optimization of the Neural Network Parameters for the Criminal Detection System
Students Name: Prodan Roman Romanovych
Qualification Level: magister
Speciality: Computer Control Systems for Moving Objects (Automobile Transport)
Institute: Institute of Computer Technologies, Automation and Metrology
Mode of Study: full
Academic Year: 2023-2024 н.р.
Language of Defence: англійська
Abstract: Recently, there has been significant interest in business and the scientific community in neural networks, particularly deep learning technologies. Their use allows for optimizing processes and solving problems much more efficiently compared to other methods. One of the current challenges is detecting criminals in residential spaces, especially given the increasing number of thefts. This master’s thesis is dedicated to investigating existing approaches to criminal detection in apartments and developing software for real-time recognition of their behavior based on neural networks. In the first chapter of our work, we carefully analyzed the history of technologies [1] aimed at criminal detection. Examining this evolutionary process in detail, we identified key stages and decisive moments that influenced the formation of modern methods. In our research, we studied various criminal detection methods, covering a wide range of technological innovations. Our analysis allowed us to gain a deeper understanding of the importance of each method and determine their effectiveness in specific conditions. We paid special attention to how each approach contributes to enhancing residential security. Additionally, we thoroughly analyzed the advantages and disadvantages of each method, taking into account their technical and ethical aspects. This provided us with an objective view of modern criminal detection technologies and their suitability for addressing current challenges in private spaces. The second chapter focuses on a new approach to criminal detection in videos using Convolutional Neural Networks (CNN) [2]. Classification and the advantages of using CNN are discussed. The network’s working stages, including parameter optimization and data augmentation, are detailed. The third chapter describes the methods and tools for implementing the technologies used. The choice of programming language (Python) and necessary libraries for the successful implementation of the neural network are determined. In the fourth chapter of our work, we extensively described the implementation of the YOLO (You Only Look Once) convolutional neural network [3] for criminal detection in videos. YOLO is notable for its unique ability to recognize and classify objects in real-time, making it particularly effective for criminal detection. We not only described the architecture of YOLO but also thoroughly discussed optimization techniques we implemented to enhance the model’s performance. This included optimizing hyperparameters such as kernel size and network depth, as well as refining the training process to ensure optimal criminal detection. Additionally, we delved into the application of data augmentation in the context of YOLO, including expanding the training dataset and using various transformations and cropping to improve the model’s performance in different conditions and ensure its robustness to various scenarios. Through the implementation of these optimization techniques and data augmentation methods, we achieved a significant improvement in the results of our YOLO model in criminal detection. These achievements define new possibilities and open perspectives for improving real-time video criminal detection systems. The fifth chapter is dedicated to the economic aspect. The development costs are calculated and compared with an existing analogue, demonstrating that the developed software solution is competitive and economically viable. In the conclusions, the importance of optimizing parameters and data augmentation for achieving high efficiency and accuracy in the real-time operation of neural networks in criminal detection is emphasized. The developed software opens up prospects for further research and development in the field of security and property protection. Keywords: apartment thefts, computer vision, deep learning, neural network, surveillance cameras. ? 1. The History and Evolution of Commercial Security Systems. URL -https://www.sutori.com/en/story/the-history-and-evolution-of-commercial- security-systems--eUmjCa8eEHnidnT4h25Rxf84 2. LeCun Y., Bengio Y., Hinton G. Deep learning. Nature. 2015. Vol. 521, no. 7553. P. 436–444. URL: https://doi.org/10.1038/nature14539 (date of access: 10.12.2023). 3. H C D. An Overview of You Only Look Once: Unified, Real-Time Object Detection. International Journal for Research in Applied Science and Engineering Technology. 2020. Vol. 8, no. 6. P. 607–609. URL: https://doi.org/10.22214/ijraset.2020.6098 (date of access: 10.12.2023).