Calculating the Effectiveness of RCNN-based Coral Reef Classification

In a prior study (Gunti & Rorissa, 2021), the coral reef classification system (Krizhevsky, Sutskever, & Hinton, 2017) encountered a massive overlap between predictions and resulted in lower precision and recall scores after submission at the Annotation and Localization Challenge (AIcrowd | ImageCLEF Coral - Annotation and Localisation | Challenges, 2022). This current study investigates the overlapping issue to improve system performance and classifies the limited number of coral substrates using the Image Conference and Labs Evaluation Forums (CLEF) coral dataset (ImageCLEFcoral | ImageCLEF / LifeCLEF - Multimedia Retrieval in CLEF, 2021- 22). The testing and evaluation were performed on the manually annotated dataset (ImageCLEFcoral | ImageCLEF / LifeCLEF - Multimedia Retrieval in CLEF, 2021-22). To improve performance, we set up the open-source objectdetection system (Yu et al., 2020) and used faster Region-Based Convolutional Neural Network (RCNN) TensorFlow models (open-source Machine Learning software library) to test the developed python script and evaluate the CLEF test Image Dataset (CoralClef) (ImageCLEFcoral | ImageCLEF / LifeCLEF - Multimedia Retrieval in CLEF, 2021-22) predictions. Our findings could assist in the refinement of the increased CNN-based retrieval systems using different models’ results to develop it more efficiently for future LifeCLEF and Mediaeval Tasks. Additionally, the study’s findings impact video retrieval, sport detection, and interview response filtering with wider significance for relevant fields.