Detecting out-of-focus plankton images

Abstract

This thesis focuses on measuring and identifying blurriness in plankton images acquired through automated plankton imaging systems, where image degradation may result from motion blur, out-of-focus acquisition, optical aberrations, water turbidity, and sensor- related issues. Since blurriness in plankton images can compromise the performance of both manual labeling and automated plankton identification, this thesis proposes a no-reference image quality assessment framework that relies solely on the input image without requiring a corresponding sharp reference image. Specifically, 29 handcrafted features are extracted from four categories: sharpness-based, statistical, frequency-based, and block-based features. These features form the basis for classifying plankton images into three categories: Blurry, In-focus, and Out-of-focus. Four tree-based classifiers, in- cluding Random Forest, Gradient Boosting, XGBoost, and a soft voting ensemble, were investigated. Experimental evaluations demonstrate that XGBoost outperforms the other models by achieving the highest cross-validation accuracy, while the ensemble model also shows strong performance. Feature analysis further indicates that edge- and gradient- based features, such as Sobel, Scharr, Laplacian, and high-energy features, contribute significantly to distinguishing different levels of image blur. Overall, the findings support the hypothesis that combining multiple handcrafted features is more effective than relying on a single blur metric. Therefore, the proposed framework provides an interpretable and effective approach for automated blur detection in plankton imagery.