Abstract
This thesis addresses the crucial need for sustainable groundfish fisheries management in the Northeastern U.S., focusing on enhancing electronic monitoring (EM) through innovative technologies. Traditionally, identifying groundfish species has been a manual and labor-intensive process. However, this research explores the use of deep learning to automate species recognition, overcoming the challenge of requiring extensive labeled data for training. The research uniquely tackles the challenge of cross-database recognition of groundfish in diverse environments, such as decks, conveyor belts, and underwater. Existing literature has not fully addressed the variations in image domains within these environments, creating a gap this thesis aims to bridge. The approach involves a transfer learning model using the YOLOv8 backbone network, known for its efficiency in real-time object detection, combined with ResNet-50, Mosaic data augmentation, class-specific anchor boxes, and the SPP-YOLO architecture, enhancing the model's performance. A critical component of this research is the development of a curated dataset featuring four key groundfish species, addressing the scarcity of underwater groundfish image datasets and enabling the training of a more robust model. This dataset is valuable for addressing the unique challenges of aquatic environments. The model demonstrates exceptional performance, achieving a mean average precision (mAP) of 94.10% in object detection and 92.14% in species classification. Beyond these quantitative results, the research significantly reduces the dependence on large volumes of labeled data for training, a notable advancement considering the resource-intensive nature of traditional methods. This achievement marks a significant step forward in pursuing sustainable fisheries management. In addition to the innovative approach and findings, a crucial aspect of this thesis is the availability of both the code and results, in the form of model weights, from all five experiments conducted. These resources are made publicly accessible, providing a foundational element for further development in the field. This availability allows for the practical application of this research in building Electronic Monitoring software, incorporating this advanced AI component for accurate fish identification. The provided model weights can also expand the model's capabilities. By introducing more complex environments into the training process, the model can be fine-tuned to recognize species in highly challenging scenarios, such as images with low light conditions or significant background clutter. This extension enhances the model's versatility and effectiveness, ensuring its applicability in real-world situations and advancing sustainable fisheries management through innovative technology.