A Video Surveillance System for Determining the Sexual Maturity of Cobia

Due to inbreeding depression, the quality of cobia seeds is decreasing; improving their quality is essential to improve the commercial value of cobia. This study proposes a system for monitoring cobia sexual maturity in culture ponds to enable fishers to identify cobia spawning times to better formulate aquaculture plans. We used a consumer-grade underwater camera to capture videos of cobia. Cobia were identified with the You Only Learn One Representation (YOLOR) object dection model. Data on detected cobia were input into a lightweight OpenPose keypoint extraction model to identify indicators of sexual maturity. The keypoints were then input into a fuzzy hyper-rectangular composite neural network (FHRCNN) to classify the cobia as being in one of four sexual maturity stages. Datasets of cobia images, sexual characteristic keypoints, and a classification scheme for sexual maturity were developed in this study. The method is lightweight and can be deployed an on edge device. The proposed model combining YOLOR, OpenPose, and FHRCNN had accuracy rates of 84.1% and 74.3% on the training and testing datasets, respectively. Ablation experiments revealed that our proposed method is superior to its state-of-the-art counterparts. We also demonstrated the superiority of the proposed method on an open dataset of facial poses.