Vinsamlegast notið þetta auðkenni þegar þið vitnið til verksins eða tengið í það: http://hdl.handle.net/1946/41552
Nematodes are a big problem in fish processing, especially in Atlantic cod. Nematodes can be hard to detect, and the repellent reactions of consumers who see them in their food can lead to a reduction in market value. Human intervention is still needed to groom and remove nematodes from fish fillets, which is a labor-intensive task. Automating the task of removing nematodes can improve the efficiency and accuracy of human labor. This thesis aims to explore the feasibility and usability of using Multispectral Imaging (MSI) for the detection of nematodes with a combination of deep learning techniques. Multispectral Imaging (MSI) data from 50 cod fillets, cut into 270 pieces, were gathered with a VideometerLab camera, which supports imaging at 19 wavelengths. Seven spectral bands were found to suffice to differentiate between fish muscle and nematodes reliably. These essential spectral bands were determined using spectral dimensionality reduction methods. Using the spectral signature of nematodes, deeply embedded nematodes at a depth of about 7 mm were detected. Due to limited data being available, we opted for a few-shot learning approach for classification. Roughly half of the images contain at least one nematode. The classification accuracy of the images went from 53% using a zero-shot learning approach to 80% accuracy when using a few-shot learning approach based on recent results in representation learning. The few-shot learning method outperformed a proven Computer Vision (CV) classification model. Using transfer learning with a pre-trained object detection model with Multispectral Imaging (MSI) data obtained 47% average precision of nematodes.
|Skemman_yfirlysing (1).pdf||38.04 kB||Lokaður||Yfirlýsing|