Multi-Scale Infrared Military Target Detection Based on 3X-FPN Feature Fusion Network

IEEE Access, Journal Year: 2023, Volume and Issue: 11, P. 141585 - 141597

Published: Jan. 1, 2023

To solve the problems of misdetection and omission infrared military targets poor detection effect in battlefield environments, an improved YOLOv4 algorithm is proposed to improve accuracy long-range target detection. First, a new 4th-scale feature extraction layer introduced enhance multi-scale sensitivity for targets. Second, TL intermediate channel realize fusion across gradient connections, 3X-FPN network structure proposed, adaptive parameters are adopted weighted balanced data accuracy. Finally, loss function established optimized model stability convergence effect. The depth separable convolution model's lightweight. experimental results vehicle class ablation show that increases by 9.85% compared with original algorithm, reduces volume 36%, its distance up 2000 m. achieves mean average precision (mAP) value 93.25% multi-military detection, which improves 12.42% mainstream meets current combat acquisition processing requirements.

Language: Английский

---

Machine learning in smart production logistics: a review of technological capabilities

International Journal of Production Research, Journal Year: 2024, Volume and Issue: unknown, P. 1 - 35

Published: July 22, 2024

Recent publications underscore the critical implications of adapting to dynamic environments for enhancing performance material and information flows.This study presents a systematic review literature that explores technological capabilities smart production logistics (SPL) when applying machine learning (ML) enhance in environments.This applies inductive theory building extends existing knowledge about SPL three ways.First, it describes role ML advancing across various dimensions, such as time, quality, sustainability, cost.Second, this demonstrates application component technologies (i.e.scanning, storing, interpreting, executing, learning) attain superior SPL.Third, outlines how manufacturing companies can cultivate effectively apply ML.In particular, introduces comprehensive framework establishes foundations SPL, thus facilitating successful integration ML, improvement capabilities.Finally, practical managers staff responsible planning execution tasks, including movement materials factories.

Language: Английский

---

ViT-DualAtt: An efficient pornographic image classification method based on Vision Transformer with dual attention

Electronic Research Archive, Journal Year: 2024, Volume and Issue: 32(12), P. 6698 - 6716

Published: Jan. 1, 2024

<p>Pornographic images not only pollute the internet environment, but also potentially harm societal values and mental health of young people. Therefore, accurately classifying filtering pornographic is crucial to maintaining safety online community. In this paper, we propose a novel image classification model named ViT-DualAtt. The adopts CNN-Transformer hierarchical structure, combining strengths Convolutional Neural Networks (CNNs) Transformers effectively capture integrate both local global features, thereby enhancing feature representation accuracy diversity. Moreover, integrates multi-head attention convolutional block mechanisms further improve accuracy. Experiments were conducted using nsfw_data_scrapper dataset publicly available on GitHub by data scientist Alexander Kim. Our results demonstrated that ViT-DualAtt achieved 97.2% ± 0.1% in tasks, outperforming current state-of-the-art (RepVGG-SimAM) 2.7%. Furthermore, achieves miss rate 1.6%, significantly reducing risk dissemination platforms.</p>

Language: Английский

---

Raw Material Flow Rate Measurement on Belt Conveyor System Using Visual Data

Applied System Innovation, Journal Year: 2023, Volume and Issue: 6(5), P. 88 - 88

Published: Sept. 30, 2023

Industries are rapidly moving toward mitigating errors and manual interventions by automating their process. The same motivation is carried out in this research which targets to study a conveyor system installed soda ash manufacturing plants. Our aim automate the determination of optimal parameters, chosen identifying flow rate materials available on belt for maintaining ratio between raw being carried. essential produce 40% pure carbon dioxide gas needed production. A visual sensor mounted used estimate materials. After selecting region interest, segmentation algorithm defined based voting-based technique segment most confident region. Moments contour features extracted passed machine learning algorithms different experiments. An in-depth analysis completed various techniques convincing results achieved final data split with best parameters using Bagging regressor. Each step process made resilient enough work challenging environment even if placed an outdoor environment. proposed solution caters current challenges serves as practical estimating material without intervention.

Language: Английский

---

Coal and Gangue Detection Networks with Compact and High-Performance Design

Sensors, Journal Year: 2024, Volume and Issue: 24(22), P. 7318 - 7318

Published: Nov. 16, 2024

The efficient separation of coal and gangue remains a critical challenge in modern mining, directly impacting energy efficiency, environmental protection, sustainable development. Current machine vision-based sorting methods face significant challenges dense scenes, where label rewriting problems severely affect model performance, particularly when are closely distributed conveyor belt images. This paper introduces CGDet (Coal Gangue Detection), novel compact convolutional neural network that addresses these through two key innovations. First, we proposed an Object Distribution Density Measurement (ODDM) method to quantitatively analyze the distribution density gangue, enabling optimal selection input feature map resolutions mitigate issues. Second, developed Relative Resolution Scale (RROSM) assess object scales, guiding design streamlined fusion structure eliminates redundant components while maintaining detection accuracy. Experimental results demonstrate effectiveness our approach; achieved superior performance with AP50 AR50 scores 96.7% 99.2% respectively, reducing parameters by 46.76%, computational cost 47.94%, inference time 31.50% compared traditional models. These improvements make suitable for real-time underground mining environments, resources limited but high accuracy is essential. Our work provides new perspective on designing yet high-performance networks scene applications.

Language: Английский

---

BCT-OFD: bridging CNN and transformer via online feature distillation for COVID-19 image recognition

International Journal of Machine Learning and Cybernetics, Journal Year: 2023, Volume and Issue: 15(6), P. 2347 - 2366

Published: Dec. 6, 2023

Language: Английский

---