A Zero Shot Detection Based Approach for Fish Species Recognition in Underwater Environments

Identification of fish species is vital for fisheries management, stock assessments, protection of endangered species, and ecosystem management. Image based surveys often deploy video cameras that are used to collect large image datasets that are reviewed by a human observer to identify species and generate a numerical count at each station. One main challenge in labeling or annotating such a large dataset is that it requires a huge amount of time, cost, and human effort. Recently, general adversarial network (GAN) based generative techniques have drawn much attention in zero-shot object detection (ZSD) because of superior performance in localizing and simultaneously recognizing objects without training a model on unseen (few target) classes. In this work, a zero-shot detection for Fish Species Recognition (ZSD-FR) in underwater environments is utilized for object detection. This approach can localize and recognize objects when the model is not trained on “unseen” classes. Generative models like GAN can be trained on data with “seen” classes for generating unseen class samples depending upon the semantics (attributes) learned from data with seen classes. The results obtained on SEAMAPD21 illustrate that the zero shot detection (ZSD) based approach of object detection can successfully transfer knowledge from seen classes to unseen classes with better detection accuracy.