Utveckling av en bildklassificeringsmodell för igenkänning av produktförpackningar i industrimiljöer : Utvärdering av domänanpassning mot standardmetoder för transfer learning

Effective waste sorting, management, and recycling are critical for sustainability and resource efficiency. Industrial recycling facilities often use conveyor belt systems, presenting opportunities for automated image-based waste classification. This thesis, conducted in collaboration with Knightec Group and Bintix, an Indian waste management startup, aims to develop and evaluate a classification model that categorizes products based on images of discarded packaging. A central challenge is the discrepancy in data quality of two available datasets. The target dataset, collected using overhead cameras on conveyor belts, contains motion blur, poor lighting, occlusions, and noisy labels. In contrast, a source dataset consists of high quality, manually labeled images captured with handheld devices in another environment. This motivates the use of unsupervised domain adaptation techniques to transfer knowledge from the clean source domain to the noisy target domain using only source labels. We explore Minimum Class Confusion (MCC) and Conditional Domain Adversarial Networks (CDAN). Ablation studies indicate that our implementations of MCC and CDAN do not clearly outperform a baseline convolutional neural network fine-tuned on the source data, achieving F1 scores between 0.264 and 0.276. Statistical tests reveal no significant differences after correcting for multiple comparisons. Incorporating a portion of labeled conveyor belt images into training results in substantial performance gains, with F1 scores between 0.714 and 0.736. Even in this setting, domain adaptation does not significantly outperform a baseline CNN. The highest performance is achieved using a standard transfer learning approach, where a ViT-B16 vision transformer trained only on conveyor belt data reaches an F1 score of up to 0.912.