Real-time Detection of Aquarium Fish Species Using YOLOv4-tiny on Raspberry Pi 4

Abstract

A version of the YOLO detection algorithm, the YOLOv4, has yet to find much use on aquatic species. Detection systems optimized for aquarium fish species are also currently lacking. This study provides a detection program for select fish species, namely the dwarf gourami, guppy, and zebrafish, using the YOLOv4-tiny detection model. The program was implemented in the Raspberry Pi 4 Model B single-board computer with an 8MP camera. The YOLOv4-tiny model was trained using images from Kaggle, FishBase, and the Global Biodiversity Information Facility, along with local images. The program was tested on live samples of the three fish species along with one irrelevant fish species, the petticoat tetra. There were three live samples of each species. Close shots for each sample were taken from the aquarium's front, left, right, and back sides, making a total of 48 images for detection. Training data and the confusion matrix from the experiment were utilized to determine the program's reliability in detecting the fish species. For the results, the trained model achieved a mAP of 97.81% during training and a global accuracy of 91.67% during the experiment. The program exhibited reliable performance across the board, achieving above 90% AP and accuracy in all classes.