Real-Time Marine Animal Detection Using Yolo-Based Deep Learning Networks in the Coral Reef Ecosystem

Abstract

Abstract. In recent years, with the advancement of marine resources and environment research, the ecological functions of reef-building coral reef ecosystems distributed in warm shallow waters of the ocean are being continuously discovered and valued by people. It is important for ecosystem protection to monitor the population of marine animals. Besides, many projects of Autonomous Underwater Vehicle (AUV) also need technology to perceive and understand environment information in real-time for better decision-making. Therefore, marine animal detection has become a challenge for researchers to study nowadays. Deep neural network models have been used to solve fish-related tasks and gained encouraging achievements, but there are still many problems in this field. In this paper, several YOLO-based methods are chosen for comparison. Experiment results indicate that these methods can recognize the marine animals in coral reef quickly and accurately. Finally, several recommendations for model improvement according to assessment results are presented.

Plot-Level Maize Early Stage Stand Counting and Spacing Detection Using Advanced Deep Learning Algorithms Based on UAV Imagery

“…Additionally, Song et al [38] used the improved YOLOX-tiny for tree height estimation based on fisheye images, and the highest relative error of the tree measurements was 4.06%. Zhong et al [39] From the visualization results, we can consider the developed method to be promising in detecting PSV in maize plots for practical use.…”

“…Additionally, Song et al [38] used the improved YOLOX-tiny for tree height estimation based on fisheye images, and the highest relative error of the tree measurements was 4.06%. Zhong et al [39] successfully adopted YOLOR to detect real-time marine animals in a coral reef ecosystem, and it achieved an AP above 0.79 in both the fish and turtle categories.…”

Enhancing aspect-based multi-labeling with ensemble learning for ethical logistics

“…However, because BR learning is predicated on the premise that labels are distinct, the subject is unaware of any potential connections between the labels, as well as the fact that the issue is dependent on the training data on more than one label because the assumption that the labels are independent is implicit in the BR learning model. Classifier Chains (CC), extensively elaborated upon, represents an innovative strategy that categorizes instances based on their association with a designated mark within a structured chain [ 11 , 33 ]. It is constructed using the BR methodology.…”

“…ReliefF multilabel submission in the work of [ 30 , 31 ] is based on probabilistic assumptions that have been previously updated. A different alteration, described in [ 32 , 33 ], took into account the nearest instances having the identical label set but distinct values. A divergent function technique with a Hamming length was presented by the author in [ 34 ].…”

High-resolution density assessment assisted by deep learning of Dendrophyllia cornigera (Lamarck, 1816) and Phakellia ventilabrum (Linnaeus, 1767) in rocky circalittoral shelf of Bay of Biscay

“…This approach entails the application of multiple layers of highly interconnected machine learning algorithms to achieve improved results from raw images ( Olden, Lawler & Poff, 2008 ; LeCun, Bengio & Hinton, 2015 ). These techniques have already achieved formidable results in different marine ecology tasks such as coral classification ( Bhandarkar, Kathirvelu & Hopkinson, 2022 ; Mahmood et al, 2017 ; Raphael et al, 2020 ), fish detection and classification ( Zhong et al, 2022 ; Siddiqui et al, 2018 ; Knausgård et al, 2022 ), and identification of diverse benthic fauna ( Abad-Uribarren et al, 2022 ; Song et al, 2022 ; Liu & Wang, 2022 ).…”