Thesis addresses the use of artificial intelligence and computer vision methods for the automation of waste management processes. The objective of this work is to create a custom dataset, train selected neural network models, evaluate the training process, and subsequently develop a software implementation of these models.

For the research, the neural networks YOLOv11 and Mask2Former were selected due to the fundamental differences in their architectural design. This enables a comparison between the convolutional architecture of YOLOv11, designed primarily for real-time detection, and Mask2Former, which represents a transformer-based architecture employing a masked attention mechanism. To compare the properties of these architectures, two datasets containing images of various types of waste were compiled. The first dataset was downloaded from the Roboflow platform and consisted of simple images featuring a single object of interest. The second dataset contained real-world images of waste containers and their contents, capturing scenes in which objects frequently overlap, thus more faithfully reflecting the conditions of potential industrial deployment.

The results indicate that Mask2Former achieves superior performance specifically on visually challenging classes, such as plastics, owing to their transparency and often highly variable and deformed shapes. However, the high segmentation accuracy of Mask2Former comes at the cost of substantially greater demands on computational memory and longer inference times, which hinders its deployment at industrial scale.
The computational requirements of YOLOv11 are considerably lower and its inference time is shorter, though this comes at the cost of noisier object detections. These results could potentially be improved by expanding the custom dataset, as the current number of images is below what is typically required for training networks of this complexity. A further finding highlighted the importance of dataset content, as models trained on images downloaded from Roboflow exhibited virtually no generalization capability.

Based on these findings, software tools were implemented separately for each architecture, enabling users to upload images or videos and perform segmentation. The software incorporates confidence threshold adjustment. This diploma thesis may serve as a foundation for further development of systems and datasets in the field of automated waste management.