Reproduzierbare Datenverteilungsänderungen zur Evaluierung der Anpassungsfähigkeit neuronaler Netze

Abstract

This thesis investigates the impact of environmental influences on object recognition in neural networks that the model did not encounter during training, as well as their compensation through fine-tuning. While pre-trained models such as SSD-MobileNetV2 are trained on large, general-purpose datasets, they often face disturbances in realworld applications that are not represented in the original training data. These include digital effects such as noise, distortions, or blurring, as well as physical influences like lens contamination or varying lighting conditions, all of which can significantly impair recognition accuracy. The aim of this work was therefore to develop a reproducible framework for systematically analyzing such influences and evaluating their effects on a neural network. For this purpose, a prototype test stand was designed and implemented, providing a fully controllable environment for image acquisition. Using a Coral board with a Coral camera, adjustable Neopixel LED rings, and plexiglass plates to simulate physical disturbances, both digital and analog environmental influences could be realistically reproduced. Based on this setup, a dataset of 9,184 images was created, covering 14 objects in 16 rotations under 14 different influences at multiple intensity levels. The dataset was annotated, converted into TFRecords, and prepared for training an SSDMobileNetV2 model. The results demonstrate that simple disturbances such as lighting changes cause only minor accuracy losses, whereas noise or contamination can reduce performance by up to 30%. Fine-tuning with targeted additional training data significantly increased the robustness of the model. In particular, the integration of noisy data proved highly beneficial, as it not only improved recognition on noisy images but also enhanced performance across other influences. The most comprehensive model, trained with all environmental factors combined, achieved the best results with an average accuracy of 87% ([email protected]:0.95). Overall, this work shows that incorporating a diverse set of realistically simulated environmental influences into the training dataset is an effective strategy to strengthen the generalization capabilities of object detection models in practical applications. This provides an important foundation for deploying AI-based recognition systems in complex real-world environments and paves the way for further research with alternative models and extended test setups.

Date of Award	2025
Original language	German (Austria)
Supervisor	Josef Langer (Supervisor)