Automated Fish Detection and Tracking System Using Pre-Trained Mask R-CNN for Ecological Biodiversity

Abstract

Introduce a new dynamic classifying algorithm in this paper to recognize and monitor fish activity to simultaneously better comprehend their synapomorphies. The pre-trained Mask Regional Convolutional Neural Network (Mask-R-CNN) is trained using a set of test models extracted from recorded video recording. The approach suggested subsequently yields well-enhanced feature vectors. The system’s automatic fish detection and tracking capabilities are improved, enhancing underwater investigation for supervising ecological biodiversity. The publicly accessible field-truth dataset assesses recall, F1-score, and classification and tracking precision. Utilizing current tracking R-CNN algorithms like Lowest Output Sum of Siamese Mask (SiamMask), Sequential Non-Maximum Suppression (Seq-NMS), and Squared Errors (MOSSE), comparative performance testing is conducted. In comparison to Siam-Mask (84%), Seq-NMS (78%) and MOSSE (75%), more than 120 of 170 specific bream were detected using the pre-trained Mask-R-CNN of the proposed algorithm (87%). This pre-trained Mask R-CNN system was used in the evaluation, and it was discovered that detection and tracking accuracy had increased significantly. This suggests that the ocean ecosystem could benefit from applying the proposed approach for management of ecology.