Object Detection and Classification based on Artificial Intelligence Methods

Abstract

Traditional computer vision methods are limited in processing complex data, such as images of wood, or white blood cells, due to limited robustness of their representation ability and susceptibility to noise. This thesis therefore proposes to apply Graph Neural Networks (GNNs) and hybrid optimization algorithms in order to improve accuracy and stability in complex image recognition. GNNs are thus used to exploit spatial and contextual relationships between image regions, helping to increase accuracy in recognition and classification. Chapter 2 explores the challenges of applying machine learning and deep learning models to real-world wood surface images and White Blood Cells (WBC) recognition, especially problems like data imbalance, image noise, or morphological ambiguity. In chapter 3 are described the theoretical foundation for Graph Attention Network (GAT) and object detection, describing spatial reasoning, attention mechanisms, and state of the art object detectors like YOLOv10 and YOLOv11. In chapter 4 is presented the evaluation system of the proposed framework in three selected application areas: Wood surface recognition using the Convolutional Neural Network (CNN) + Label Propagation Particle Swarm Optimization (LPMPSO) + Graph Attention Network Variational Autoencoder (GATVAE) model as the first one, WBC classification with the YOLOv10 + Multi-Hop Graph Attention Network (MHGAT) model as the second one, and finally the YOLOv11 + GAT + Transformer-based head as the last application area. In chapter 5 are summarized the results obtained from the proposed models and the effectiveness in two practical applications is evaluated. In this chapter, the contribution of the thesis in overcoming the limitations of traditional methods and proposes directions for future applications is highlighted. In summary, research conducted in this thesis improves the efficiency of wood surface defects classification and white blood cell classification. Combination of GNN with deep learning and modern optimization improves the accuracy and adaptability in real-world environments and opens up new research potential for graph data-based systems.