Generic placeholder image

Recent Patents on Engineering

Editor-in-Chief

ISSN (Print): 1872-2121
ISSN (Online): 2212-4047

Review Article

YOLOv5-based Enhanced Underwater Seaweed Detection Using Open-Source Datasets

In Press, (this is not the final "Version of Record"). Available online 30 May, 2024
Author(s): Prabhakar G*, Nirmal Kannan V, Selvaperumal S and Baranitharan B
Published on: 30 May, 2024

Article ID: e300524230521

DOI: 10.2174/0118722121303581240527070227

Price: $95

Abstract

Seaweeds, macroscopic algae known for their diverse applications across industries such as agriculture, biomedicine, and personal care, pose challenges in underwater detection due to the lack of suitable image datasets. Underwater photography is often hindered by light fluctuations and artificial light-induced haze and noise, complicating accurate object recognition. To address these challenges, this study leverages openly available open-source datasets such as Algae detection Dataset, seaweeddetection- test Dataset, and Fish detection Dataset. From these datasets, seaweed-spotted images, along with sea creatures and coral reefs, are segregated for employment in this research work. Employing a supervised learning algorithm facilitated by Roboflow software and the YOLOv5 architecture, the method achieves a commendable accuracy of 82.3% in identifying seaweed presence in underwater images. The investigation not only highlights the potential of seaweeds as renewable resources but also underscores the intricate nature of underwater detection. Moreover, the approach offers a promising avenue for refining Underwater Object Detection (UOD) algorithms, benefiting not only seaweed detection but also the identification of various marine species. Additionally, acknowledgment is made of the diversity of patented underwater object detection technologies, each with its own set of advantages, disadvantages, and application scenarios. To address the challenge of detecting seaweeds underwater, a supervised learning algorithm was developed utilizing the Roboflow software and YOLOv5 architecture. The selection of the dataset was guided by the need for a diverse representation of underwater scenes containing seaweeds, sea creatures, and coral reefs. Openly available open-source datasets, including the Algae detection Dataset, seaweed-detection-test Dataset, and Fish detection Dataset, were chosen based on their coverage of these underwater environments. From these datasets, seaweedspotted images, alongside images of sea creatures and coral reefs, were segregated to form a comprehensive training dataset. This selection was made to ensure the inclusion of diverse underwater scenes containing seaweeds, thus enhancing the robustness of the model. Data augmentation techniques, including rotation, scaling, and flipping, were applied to increase dataset diversity and robustness. During the training process, model parameters were iteratively optimized to achieve accurate seaweed detection. This involved fine-tuning the YOLOv5 architecture and adjusting training parameters based on the characteristics of the selected dataset. By utilizing openly available datasets and employing data augmentation techniques, the methodology ensures the inclusion of diverse underwater scenes containing seaweeds, effectively addressing the challenges associated with underwater detection. The study evaluated the performance of the YOLOv5 algorithm in detecting underwater objects, focusing on Fish, Seaweed, and Starfish classes. Seaweed detection demonstrated high precision (0.925) and mAP50 (0.823), with an accuracy rate of 82.3% across diverse underwater images. Fish detection showed moderate precision (0.769) and mAP50 (0.512), while Starfish detection exhibited exceptional accuracy (Precision: 0.872, Recall: 1, mAP50: 0.995). These findings emphasize the algorithm's effectiveness in identifying underwater objects, particularly seaweed and starfish, highlighting its potential for marine species identification tasks. In conclusion, the YOLOv5 algorithm demonstrated remarkable accuracy in identifying seaweed, achieving an initial accuracy of 82.3%. Through subsequent retraining, both classification and detection accuracy significantly improved, reaching 77.7% and 99.5%, respectively. Additionally, the mAP50 value for seaweed also reaches 82.3%, indicating robust performance. However, for fish species, the accuracy is recorded at 51.2%, with detection accuracy also reaching 99.5%. These findings underscore the versatility and effectiveness of the YOLOv5 algorithm in underwater object detection tasks, highlighting its potential for enhancing marine species identification and ecological research.


Rights & Permissions Print Cite
© 2025 Bentham Science Publishers | Privacy Policy