Published: July 9, 2024
Language: Английский
Food and Bioprocess Technology, Journal Year: 2023, Volume and Issue: 17(4), P. 828 - 849
Published: July 8, 2023
Language: Английский
Citations
9
Frontiers in Marine Science, Journal Year: 2023, Volume and Issue: 10
Published: June 5, 2023
Image-based machine learning methods are becoming among the most widely-used forms of data analysis across science, technology, engineering, and industry. These powerful because they can rapidly automatically extract rich contextual spatial information from images, a process that has historically required large amount human labor. A wide range recent scientific applications have demonstrated potential these to change how researchers study ocean. However, despite their promise, tools still under-exploited in many domains including species environmental monitoring, biodiversity surveys, fisheries abundance size estimation, rare event detection, animal behavior, citizen science. Our objective this article is provide an approachable, end-to-end guide help apply image-based effectively own research problems. Using case study, we describe prepare data, train deploy models, overcome common issues cause models underperform. Importantly, discuss diagnose problems poor model performance on new imagery build robust vastly accelerate acquisition marine realm. Code perform analyses provided at https://github.com/heinsense2/AIO_CaseStudy .
Language: Английский
Citations
8
Sensors, Journal Year: 2024, Volume and Issue: 24(8), P. 2430 - 2430
Published: April 10, 2024
Long-term, automated fish detection provides invaluable data for deep-sea aquaculture, which is crucial safe and efficient seawater aquafarming. In this paper, we used an infrared camera installed on a truss-structure net cage to collect images, were subsequently labeled establish dataset. Comparison experiments with our dataset based Faster R-CNN as the basic objection framework conducted explore how different backbone networks network improvement modules influenced performances. Furthermore, also experimented effects of learning rates, feature extraction layers, augmentation strategies. Our results showed that EfficientNetB0 FPN module was most competitive dataset, since it took significantly shorter time while maintaining high AP50 value 0.85, compared best 0.86 being achieved by combination VGG16 all plus augmentation. Overall, work has verified effectiveness deep learning-based object methods provided insights into subsequent improvements.
Language: Английский
Citations
3
Journal of Marine Science and Engineering, Journal Year: 2024, Volume and Issue: 12(7), P. 1132 - 1132
Published: July 5, 2024
The real-time classification of fish feeding behavior plays a crucial role in aquaculture, which is closely related to cost and environmental preservation. In this paper, Fish Feeding Intensity model based on the improved Vision Transformer (CFFI-Vit) proposed, capable quantifying behaviors rainbow trout (Oncorhynchus mykiss) into three intensities: strong, moderate, weak. process outlined as follows: firstly, we obtained 2685 raw images from recorded videos classified them categories: Secondly, number transformer encoder blocks internal structure ViT was reduced 12 4, can greatly reduce computational load model, facilitating its deployment mobile devices. And finally, residual module added head ViT, enhancing model’s ability extract features. proposed CFFI-Vit has 5.81 G (Giga) Floating Point Operations per Second (FLOPs). Compared original it reduces demands by 65.54% improves accuracy validation set 5.4 percentage points. On test set, achieves precision, recall, F1 score 93.47%, 93.44%, 93.42%, respectively. Additionally, compared state-of-the-art models such ResNet34, MobileNetv2, VGG16, GoogLeNet, higher 6.87, 8.43, 7.03, 5.65 points, Therefore, achieve while significantly reducing demands. This provides foundation for deploying lightweight deep network edge devices with limited hardware capabilities.
Language: Английский
Citations
2
Journal of Marine Science and Engineering, Journal Year: 2024, Volume and Issue: 12(10), P. 1823 - 1823
Published: Oct. 12, 2024
Real-time estimation of fish biomass plays a crucial role in real-world fishery production, as it helps formulate feeding strategies and other management decisions. In this paper, dense counting network called Swin-CSRNet is proposed. Specifically, the VGG16 layer front-end replaced with Swin transformer to extract image features more efficiently. Additionally, squeeze-and-excitation (SE) module introduced enhance feature representation by dynamically adjusting importance each channel through “squeeze” “excitation”, making extracted focused effective. Finally, multi-scale fusion (MSF) added after back-end fully utilize information, enhancing model’s ability capture details. The experiment demonstrates that achieved excellent results MAE, RMSE, MAPE correlation coefficient R2 11.22, 15.32, 5.18%, 0.954, respectively. Meanwhile, compared original network, parameter size computational complexity were reduced 70.17% 79.05%, Therefore, proposed method not only counts number higher speed accuracy but also contributes advancing automation aquaculture.
Language: Английский
Citations
2
3C Tecnología_Glosas de innovación aplicadas a la pyme, Journal Year: 2023, Volume and Issue: 12(2)
Published: June 25, 2023
The successful promotion of an academic at institution higher learning is affected, to a large degree, by the publication record applicant.This usually updated in resume, portfolio, or online database, such as Google Scholar, Research Gate, LinkedIn.The purpose this article present metric comparison between Scholar and Gate for rated scientists who are employed universities technology South Africa.This may help establish notable similarities differences from specific identify which platform they prefer maintain their records.A snapshot quantitative study used where total number citations, h-index values, scores were collected analyzed.Results indicate that has highest authors six technology, with recording values these authors.Only 134 out 181 (sample size) records on both databases.It recommended researchers education use least one database publications, thereby enhancing visibility research done university enabling more valid performance achieved each researcher.
Language: Английский
Citations
5
Journal of Food Composition and Analysis, Journal Year: 2023, Volume and Issue: 125, P. 105746 - 105746
Published: Oct. 11, 2023
Language: Английский
Citations
4
Animals, Journal Year: 2024, Volume and Issue: 14(9), P. 1368 - 1368
Published: May 1, 2024
With the rapid development of turtle breeding industry in China, demand for automated sorting is increasing. The automatic Chinese softshell turtles mainly consists three parts: visual recognition, weight prediction, and individual sorting. This paper focuses on two aspects, i.e., recognition a novel method object detection prediction proposed. In process, computer vision technology used to estimate classify them by weight. For body parts turtles, color space model proposed this separate from background effectively. By applying multiple linear regression analysis modeling, relationship between morphological parameters obtained, which can be well. An improved deep learning network extract features plastron carapace achieving excellent results. mAP reached 96.23%, meet requirements accurate identification turtles.
Language: Английский
Citations
1
3C Tecnología_Glosas de innovación aplicadas a la pyme, Journal Year: 2023, Volume and Issue: unknown, P. 347 - 362
Published: June 25, 2023
The problem of chip defects has always existed in industrial production, and since there are more environmental problems caused by defects, people have attached greater importance to the identification detection defects. Pursuant ecological process this paper uses machine vision technology detect high-performance phase noise measurement chips. results suggest that accuracy for reaches up 98%. production volume organic waste gas decreases from 5968.0t/a 4000t/a. yield wastewater 5496m3/d 4600m3/d. amount solid reduces 8000t/a 6500t/a. aforementioned data all confirm advantages automation, high efficiency, defect And also, improving discharge gas, wastewater, is reduced, thereupon environment ameliorated.
Language: Английский
Citations
3
Research Square (Research Square), Journal Year: 2023, Volume and Issue: unknown
Published: April 19, 2023
Abstract This study proposes a novel framework for fish species classification that combines FRCNN (Faster Region-based Convolutional Neural Network), VGG16 (Visual Geometry Group 16), and SPPNet (Spatial Pyramid Pooling network). The proposed FRCNN-VGG16-SPPNet the strengths of FRCNN's fast object detection localization, VGG16's convenient transfer learning performance, SPPNet's image processing flexibility robustness in handling input images any size. First, is used to detect extract target objects from containing multiple objects. Subsequently, photos various at different scales are fed into VGG16-SPPNet, which performs basic feature extraction using theory. further processes by performing pooling operations scales. Finally, identifies important features perform classification. achieves higher accuracy compared traditional single models, particularly classifying sizes, with an rate 0.9318, 26% than models. efficient, convenient, reliable, robust has potential applications recognition
Language: Английский
Citations
2