Automated fish detection and tracking system using pre-trained Mask R-CNN for Ecological biodiversity

In this paper, propose a new dynamic classifying algorithm which supports fully automatic fish detection and tracking system to identify fish species and to track fish activities for understanding synapomorphies characteristic simultaneously. The pre-trained Mask Regional Convolutional Neural Network (Mask-R-CNN) is involved for having well-enhanced feature vectors derived after undergone with number of test samples taken from captured video footage. Hence, the proposed framework is noticed as pre-trained Mask-R-CNN. It improves the system function of automatic fish detection and tracking to enhance the underwater surveillance for monitoring ecological biodiversity. The available ground–truth dataset is used to evaluate the system precisian, F1-score, and recall in terms of classifying and tracking mechanism. The comparative analysis is made with existing tracking R-CNN algorithms such as minimum output sum of squared errors (MOSSE), sequential non-maximum suppression (Seq-NMS) and Siamese mask (SiamMask). Simulation results conveys that the proposed algorithm support effective fish detection (i.e. around 120 out of 170 individual bream) and accuracy of the pre-trained Mask-R-CNN (87%) as compared with MOSSE (75%), Seq-NMS (78%), and SiamMask (84%) respectively. Thus, the evaluation result shows, the proposed pre-trained Mask-R-CNN achieves reasonable improvement in accuracy (detection and tracking) which give potential benefits to ocean ecosystem for establishing the ecological management.