Landmarking Technique for Improving YOLOv4 Fish Recognition in Various Background Conditions

Abstract

The detection and classification of fish is a prevalent and fascinating area of study.Numerous researchers develop skills in fish recognition in both aquatic and non-aquatic environments, which is beneficial for population control and the aquaculture industry, respectively.Rarely is research conducted to optimize the recognition of fish with diverse backgrounds.This paper proposes a method for fish recognition that uses the landmarking technique to optimize YOLO version 4 to detect and classify fish with varying background conditions, making it applicable for both underwater and terrestrial fish recognition.The proposed procedure was evaluated on the Bringham Young University (BYU) dataset containing four different fish species.The final test results indicate that the detection accuracy had reached 96.60% with an average confidence score of 99.67% at the 60% threshold.The outcome is 4,94% better than the most common traditional labeling method.