Design and Implementation of Chaetodontidae Fish Identification Algorithms with Deep Learning Methods

Chaetodontidae fish is one of the indicators of coral reef health, so it is necessary to observe its abundance. Autonomous Underwater Vehicle (AUV) is expected to be able to identify fish in the waters faster than conventional methods because it can automatically recognize objects that are being recorded by implementing deep learning algorithms in it. This research aims to compare the performance of three algorithms (SSD-MobileNet, Faster-RCNN, and TinyYOLO) and determine the appropriate algorithm to be implemented to the AUV. Models with the highest to lowest accuracy and precision are Faster-RCNN, SSD-MobileNet, and TinyYOLO. Models with the highest to lowest computing speed are SSD- MobileNet, TinyYOLO, and Faster-RCNN. SSD-MobileNet is stated to have the best performance with a mAP value of about 84.47% and a framerate of about 3.08 fps. The computation speedof SSD-MobileNet, when implemented on the Raspberry Pi, is around 18.02 fps with the addition of the Coral USB Accelerator. This allows the fish identification using the AUV to be accuratein real-time.