PT Unknown AU Schröder, S TI Improvement of Semi-Automatic Image Classification for Plankton Research using Machine Learning SE Kiel Computer Science Series PY 2011 VL 2024/4 PU Christian-Albrechts-Universität zu Kiel DI 10.21941/kcss/2024/4 WP https://macau.uni-kiel.de/receive/macau_mods_00005500 LA en DE Plankton Image Classification; Low-shot Learning; MorphoCluster; Image Annotation; Clustering; Feature Extraction; Supervised Learning; Self-supervised Learning; Transfer Learning SN 2194-6639 AB This cumulative dissertation is concerned with investigating and developing approaches aimed at enhancing the efficiency and precision of image classification in the domain of biological oceanography. The first publication investigates a low-shot learning approach tailored for plankton image data. This research uses a low-shot technique called "weight imprinting" to address the constraints posed by limited training data. Due to its quick retraining capabilities, this approach holds potential for evolving datasets and adding new classes efficiently. The second publication introduces MorphoCluster, a novel image annotation tool designed to address the limitations of traditional classifier-based annotation approaches. Through unsupervised clustering and similarity search mechanisms, MorphoCluster enables rapid computer-assisted image classification. We demonstrate its effectiveness using a example data set of 1.2 million plankton images. The third publication assesses feature extractors and clustering algorithms within the MorphoCluster framework using simulated user behavior. Exploring various training regimes and algorithmic choices, this research evaluates supervised, self-supervised, and transfer representation learning approaches. Additionally, two novel clustering algorithms are introduced and compared against established methods, shedding light on their effectiveness in the context of image annotation. Collectively, these articles test proven approaches and contribute novel methodologies and comprehensive assessments aimed at enhancing image annotation in biological oceanography. The findings offer potential for improving efficiency and accuracy in image classification and, thereby, facilitating more accurate and more comprehensive analyses in the study of marine ecosystems. PI Kiel ER