Analysis of deep learning algorithms for prawn aquaculture in a challenging environment

Abstract

Deep Learning plays a vital role in various domains, spanning surveillance, agriculture, and healthcare. However, the current landscape lacks comprehensive comparative studies, underscoring the needs to explore various convolutional neural networks (CNNs) for aquaculture industries. This study conducts a comprehensive comparative analysis on state-of-the-art object detection models to identify the most suitable approach for detecting freshwater prawns, specifically in challenging turbid and murky water conditions as there is a lack of comparison studies for underwater images of freshwater prawns. The evaluated models, including YOLOv8, Faster R-CNN (FRCNN), Single Shot Detector (SSD), and EfficientDet, are assessed based on the following metrics namely inference time, accuracy, image cropping size, and model memory footprint. Using a dataset comprising 2976 manually collected images of freshwater prawns, augmented to a total of 7148 images, our results demonstrate that YOLOv8 outperforms its counterparts, achieving an impressive mean Average Precision (mAP) accuracy of 83.71% with an inference time of 32.9 milliseconds. This positions YOLOv8 as a dependable choice for real-time object detection in challenging underwater environments and contributes to addressing the limited research comparing YOLOv8 with other popular object detection algorithms, thereby aiding model selection for similar object detection applications.