Detekce objektů v obrazech s využitím difuzních neuronových sítí

Abstract

This diploma thesis deals with the application of generative diffusion models in computer vision tasks, focusing on object and anomaly detection. The theoretical part analyzes the limitations of traditional discriminative architectures (CNN, DETR) and details the principles of iterative denoising, particularly the modern Noise-to-Box paradigm. In the practical part, a custom DiffusionDet architecture is designed and fully implemented using the PyTorch framework. This model is subsequently trained and optimized using stochastic data augmentations on two specific datasets: vehicles (Baltrip) and traffic signs (Roadsigns). Through extensive ablation experiments, the impact of the Transformer decoder’s structural capacity on learning stability, localization accuracy (mAP), and computational inference throughput (FPS) is investigated. Furthermore, the thesis extends into the domain of unsupervised anomaly detection, practically demonstrating the reconstructive capabilities of DDPM models on medical data and analyzing their potential for industrial quality control. The result of the thesis is an evaluation of the applicability of diffusion networks for real-time detection.