Fish Detection using Dual Branch Image Enhanced Pipeline with YOLOv8

Fish species identification plays a crucial role in the aquaculture industry. Current methods such as using sensors, acoustic devices and density map regression, face multiple challenges related to costs, computational complexity and corrosion or malfunctioning of devices in underwater environment. To address these limitations we use YOLOv8(You Only Look Once) for the identification of fish species using visual perspective through the camera. The model is trained using a dataset of fish aquarium images to provide ground truth labels for training and evaluation purposes. Two species were observed in this research: Iridescent Shark Catfish (Pangasianodon hypophthalmus) and Goldfish (Carassius auratus). Our study achieved a notable breakthrough in identifying all the two species within a comprehensive dataset comprising 173 images. The top-performing model demonstrated exceptional accuracy on the test dataset with splitting ratios of 70-20-10 for training, validation, and testing, respectively. The custom model attained a mAP which is mean Average Precision, of 92% at an IoU threshold of 0.5, underscoring its impressive capabilities to detect across all classes. To remove the turbidity, a dual branch image enhancement pipeline is used with yolov8, which gives the best confidence score of 66% in detecting the fish species.