Inconsistency-based active learning with adaptive pseudo-labeling for fish species identification DOI
M M Nabi, Chiranjibi Shah, Simegnew Yihunie Alaba

et al.

Published: June 6, 2024

The deep neural network has found widespread application in object detection due to its high accuracy. However, performance typically depends on the availability of a substantial volume accurately labeled data. Several active learning approaches have been proposed reduce labeling dependency based confidence detector. Nevertheless, these tend exhibit biases toward high-performing classes, resulting datasets that do not adequately represent testing In this study, we introduce comprehensive framework for considers both uncertainty and robustness detector, ensuring superior across all classes. robustness-based score is calculated using consistency between an image augmented version. Additionally, leverage pseudo-labeling mitigate potential distribution drift enhance model performance. To address challenge setting threshold, adaptive threshold mechanism. This adaptability crucial, as fixed can negatively impact performance, particularly low-performing classes or during initial stages training. For our experiment, employ Southeast Area Monitoring Assessment Program Dataset 2021 (SEAMAPD21), comprising 130 fish species with 28,328 samples. results show outperforms state-of-the-art method significantly reduces annotation cost. Furthermore, benchmark model's against public dataset (PASCAL VOC07), showcasing effectiveness comparison existing methods.

Language: Английский

MuLA-GAN: Multi-Level Attention GAN for Enhanced Underwater Visibility DOI Creative Commons
Ahsan Baidar Bakht,

Zikai Jia,

Muhayy Ud Din

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 81, P. 102631 - 102631

Published: May 11, 2024

The underwater environment presents unique challenges (color distortions, reduced contrast, blurriness) hindering accurate analysis. This work introduces MuLA-GAN, a novel approach leveraging Generative Adversarial Networks (GANs) and specifically adapted Multi-Level Attention for comprehensive image enhancement. MuLA-GAN integrates within the GAN architecture to prioritize learning discriminative features crucial precise restoration. These relevant encompass information on local details regions leveraged by spatial attention at various scales across entire captured multi-level attention. allows identify enhance objects, textures, edges obscured distortions while also reconstructing more visually clear representation of scene analyzing low-level like as well high-level object shapes global information. By selectively focusing these features, excels capturing preserving intricate in imagery, which is essential marine research, exploration, resource management applications. Extensive evaluations diverse datasets (UIEB test, UIEB challenge, U45, UCCS) demonstrate MuLA-GAN's superior performance compared existing methods. Additionally, specialized bio-fouling aquaculture dataset confirms model's robustness challenging environments. On test dataset, achieves exceptional Peak Signal-to-Noise Ratio (PSNR) (25.59) Structural Similarity Index (SSIM) (0.893) scores, surpassing Water-Net (24.36 PSNR, 0.885 SSIM). addresses significant research gap enhancement demonstrating effectiveness combining GANs with mechanisms. tailored offers framework restoring quality, providing valuable insights source code publicly available GitHub https://github.com/AhsanBaidar/MuLA_GAN.git

Language: Английский

Citations

6
Multi-species identification and number counting of fish passing through fishway at hydropower stations with LigTraNet DOI Creative Commons
Jianyuan Li,

Chunna Liu,

Luhai Wang

et al.

Ecological Informatics, Journal Year: 2024, Volume and Issue: 82, P. 102704 - 102704

Published: June 26, 2024

Fishway monitoring can verify the effectiveness of fishway, optimise operation mode, and achieve scientific management fishway operations. Traditional approaches, hindered by their inefficiency substantial disruption fish, are ill-suited for long-term surveillance; thus, employing video coupled with object detection technology presents an alternative or complementary solution. However, challenges such as constrained computational capacity onsite equipment in fishways, complexities involved model deployment, sluggish pace significant hurdles. In this study, utilising YOLOv8n a benchmark, we engineered cross-stage partial module single convolution (C1) to replace existing C2f aim enhancing performance. We replaced conventional 2D convolutions bottleneck configuration depthwise separable integrated SimAM extract detailed characteristics fish species. By amalgamating LigObNet DeepSORT algorithm, established LigTraNet, which is designed enable precise tracking, identification, counting individual fish. The results showed that exhibited lowest complexity fastest speed underwater among similar recognition models. Compared benchmark model, there were reductions 8.9% network layers, 40.5% parameter count, 39.3% memory footprint, 35.8% giga floating-point operations 38.1% improvement inference speed. LigTraNet achieved total count accuracy rate 91.8%, demonstrating superior species quantification capabilities over other models minimal resource usage rapid capabilities, thus offering enhanced practicality deployment on devices real-world engineering contexts. This represents departure from traditional manual methods assessing effectiveness, revolutionising aquatic ecological tools methodologies fostering collaborative advancement water project conservation.

Language: Английский

Citations

5
Neural Network for Underwater Fish Image Segmentation Using an Enhanced Feature Pyramid Convolutional Architecture DOI Creative Commons
Guang Yang, Junyi Yang,

Wenyao Fan

et al.

Journal of Marine Science and Engineering, Journal Year: 2025, Volume and Issue: 13(2), P. 238 - 238

Published: Jan. 26, 2025

Underwater fish image segmentation is a crucial technique in marine monitoring. However, typical underwater images often suffer from issues such as color distortion, low contrast, and blurriness, primarily due to the complex dynamic nature of environment. To enhance accuracy segmentation, this paper introduces an innovative neural network model that combines attention mechanism with feature pyramid module. After backbone processes input through convolution, data pass enhanced module, where it iteratively processed by multiple weighted branches. Unlike conventional methods, multi-scale extraction module we designed not only improves high-level semantic features but also optimizes distribution low-level shape weights synergistic interactions branches, all while preserving inherent properties image. This novel architecture significantly boosts accuracy, offering new solution for tasks. further model’s robustness, Mix-up CutMix augmentation techniques were employed. The was validated using Fish4Knowledge dataset, experimental results demonstrate achieves Mean Intersection over Union (MIoU) 95.1%, improvements 1.3%, 1.5%, 1.7% MIoU, Pixel Accuracy (PA), F1 score, respectively, compared traditional methods. Additionally, real dataset captured deep-sea environments constructed verify practical applicability proposed algorithm.

Language: Английский

Citations

0
DeepFins: Capturing dynamics in underwater videos for fish detection DOI Creative Commons
Ahsan Jalal, Ahmad Salman, Ajmal Mian

et al.

Ecological Informatics, Journal Year: 2025, Volume and Issue: unknown, P. 103013 - 103013

Published: Jan. 1, 2025

Language: Английский

Citations

0
Optimizing feature map matching for marine benthic organism detection DOI
Xinzhi Li, Yong Liu, Peng Yan

et al.

The Visual Computer, Journal Year: 2025, Volume and Issue: unknown

Published: March 5, 2025

Language: Английский

Citations

0
How to track and segment fish without human annotations: a self-supervised deep learning approach DOI Creative Commons
Alzayat Saleh, Marcus Sheaves, Dean R. Jerry

et al.

Pattern Analysis and Applications, Journal Year: 2024, Volume and Issue: 27(1)

Published: Feb. 23, 2024

Abstract Tracking fish movements and sizes of is crucial to understanding their ecology behaviour. Knowing where migrate, how they interact with environment, size affects behaviour can help ecologists develop more effective conservation management strategies protect populations habitats. Deep learning a promising tool analyse from underwater videos. However, training deep neural networks (DNNs) for tracking segmentation requires high-quality labels, which are expensive obtain. We propose an alternative unsupervised approach that relies on spatial temporal variations in video data generate noisy pseudo-ground-truth labels. train multi-task DNN using these pseudo-labels. Our framework consists three stages: (1) optical flow model generates the pseudo-labels consistency between frames, (2) self-supervised refines incrementally, (3) network uses refined labels training. Consequently, we perform extensive experiments validate our method public datasets demonstrate its effectiveness annotation segmentation. also evaluate robustness different imaging conditions discuss limitations.

Language: Английский

Citations

2
Novel use of deep neural networks on photographic identification of epaulette sharks (Hemiscyllium ocellatum) across life stages DOI Creative Commons
M. Lonati, Mohammad Jahanbakht,

Danielle Atkins

et al.

Journal of Fish Biology, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 10, 2024

Photographic identification (photo ID) is an established method that used to count animals and track individuals' movements. This performs well with some species of elasmobranchs (i.e., sharks, skates, rays) where individuals have distinctive skin patterns. However, the unique patterns for ID must be stable through time allow re-identification in future sampling events. More recently, artificial intelligence (AI) models substantially decreased labor-intensive process matching photos extensive photo libraries increased reliability ID. Here, AI are first identify epaulette sharks (Hemiscyllium ocellatum) at different life stages approximately 2 years. An model was developed assess compare human-classified juvenile neonate sharks. The also tested persistence adult Results indicate immature unreliable pattern identification, using both human approaches, due plasticity these subadult growth forms. Mature maintain their can identified by 86% accuracy. approach outlined this study has potential validating stability time; however, testing on wild populations long-term datasets needed. study's novel deep neural network development strategy offers a streamlined accessible framework generating reliable from small data set, without requiring high-performance computing. Since many studies commence limited resources, presents practical solutions such constraints. Overall, address challenges associated sets application shark identification.

Language: Английский

Citations

1
Semi-supervised learning advances species recognition for aquatic biodiversity monitoring DOI Creative Commons
Dongliang Ma,

Jine Wei,

Likai Zhu

et al.

Frontiers in Marine Science, Journal Year: 2024, Volume and Issue: 11

Published: May 28, 2024

Aquatic biodiversity monitoring relies on species recognition from images. While deep learning (DL) streamlines the process, performance of these method is closely linked to large-scale labeled datasets, necessitating manual processing with expert knowledge and consume substantial time, labor, financial resources. Semi-supervised (SSL) offers a promising avenue improve DL models by utilizing extensive unlabeled samples. However, complex collection environments long-tailed class imbalance aquatic make SSL difficult implement effectively. To address challenges in within scheme, we propose Wavelet Fusion Network Consistency Equilibrium Loss function. The former mitigates influence data environment fusing image information at different frequencies decomposed through wavelet transform. latter improves scheme refining consistency loss function adaptively adjusting margin for each class. Extensive experiments are conducted FishNet dataset. As expected, our existing up 9.34% overall classification accuracy. With accumulation data, improved limited shows potential advance conservation.

Language: Английский

Citations

0
Marine Resources: Identification, Restoring, and Monitoring of Fisheries Food Resources Using Deep Learning and Image Processing DOI

N. Nasurudeen Ahamed,

Amreen Ayesha

Published: Jan. 1, 2024

Language: Английский

Citations

0
Fisheries Management with Deep Learning-Based Fish Species Detection: A Sustainable Approach DOI

Nemi Rishi,

Akhil Kumar,

Richa Golash

et al.

Lecture notes in networks and systems, Journal Year: 2024, Volume and Issue: unknown, P. 359 - 369

Published: Dec. 4, 2024

Language: Английский

Citations

0