Applying Image Recognition to Enhance Fisheries Management Capabilities

Marine fisheries populations have a large impact on the U.S. economy – from commercial fishing to coastal communities. Overfishing, barriers to migration, and other forms of human activity may impact spawning patterns of these species. Therefore, it is necessary to monitor these populations to maintain sustainable resources, healthy oceans, and marine life. Federal and state agencies deploy camera equipment to monitor fisheries populations; employees then manually count the number of specimens in the gathered images – a process that is both time- and labor-intensive. We propose an alternative approach. Through the application of deep learning-based image recognition, identification of target species in photographic data can be automated. Current state-of-the-art image recognition relies on Convolutional Neural Networks (CNNs) to achieve learning and recognition. CNNs loosely represent biological neural networks: each neuron, or layer, accomplishes a specific task, such as edge detection. Such algorithms can be adopted to enhance the capabilities of fisheries management in monitoring fisheries populations. This research introduces a new set of training data and explores the accuracy of image recognition algorithms, such as You Only Look Once (YOLO) v2: Real-Time Object Detection, in detecting target species, including alewife herring (Alosa pseudoharengus), blueback herring (Alosa aestivalis), Atlantic sea scallop (Placopecten magellanicus), flatfish such as flounder (Pleuronectiformes), skates (Rajidae), and various round fish species. Factors pertaining to accuracy, such as quality of annotations and number of training iterations, are also explored. Results suggest this approach is viable; average recall, a metric for accuracy, reaches as high as 93% in the precision-recall curves obtained. Therefore, fisheries management can preserve resources by applying image recognition in its stock assessments of marine fisheries populations.