Discrepant Semantic Diffusion Boosts Transfer Learning Robustness DOI Open Access
Yajun Gao, Shihao Bai, Xiaowei Zhao

et al.

Electronics, Journal Year: 2023, Volume and Issue: 12(24), P. 5027 - 5027

Published: Dec. 16, 2023

Transfer learning could improve the robustness and generalization of model, reducing potential privacy security risks. It operates by fine-tuning a pre-trained model on downstream datasets. This process not only enhances model’s capacity to acquire generalizable features but also ensures an effective alignment between upstream knowledge domains. can effectively speed up convergence when adapting novel tasks, thereby leading efficient conservation both data computational resources. However, existing methods often neglect discrepant downstream–upstream connections. Instead, they rigidly preserve information without adequate regularization semantic discrepancy. Consequently, this results in weak generalization, issues with collapsed classification, overall inferior performance. The main reason lies connection due mismatched granularity. Therefore, we propose diffusion method for transfer learning, which adjust granularity alleviate classification problem Specifically, proposed framework consists Prior-Guided Diffusion pre-training fine-tuning. Firstly, aims empower semantic-diffusion ability. is achieved through prior, consequently provides more robust classification. Secondly, focuses encouraging diffusion. Its design intends avoid unwanted centralization, causes Furthermore, it constrained discrepancy, serving elevate discrimination capabilities. Extensive experiments eight prevalent datasets confirm that our outperform number state-of-the-art approaches, especially fine-grained or dissimilar (e.g., 3.75% improvement Cars dataset 1.79% SUN under few-shot setting 15% data). sparsity caused protection successfully validate method’s effectiveness field artificial intelligence security.

Language: Английский

Hybrid Population Based Training–ResNet Framework for Traffic-Related PM2.5 Concentration Classification DOI Creative Commons
Afaq Khattak, Badr T. Alsulami, Caroline Mongina Matara

et al.

Atmosphere, Journal Year: 2025, Volume and Issue: 16(3), P. 303 - 303

Published: March 5, 2025

Traffic emissions serve as one of the most significant sources atmospheric PM2.5 pollution in developing countries, driven by prevalence aging vehicle fleets and inadequacy regulatory frameworks to mitigate effectively. This study presents a Hybrid Population-Based Training (PBT)–ResNet framework for classifying traffic-related levels into hazardous exposure (HE) acceptable (AE), based on World Health Organization (WHO) guidelines. The integrates ResNet architectures (ResNet18, ResNet34, ResNet50) with PBT-driven hyperparameter optimization, using data from Open-Seneca sensors along Nairobi Expressway, combined meteorological traffic data. First, analysis showed that PBT-tuned ResNet34 was effective model, achieving precision (0.988), recall (0.971), F1-Score (0.979), Matthews Correlation Coefficient (MCC) 0.904, Geometric Mean (G-Mean) 0.962, Balanced Accuracy (BA) outperforming alternative models, including ResNet18, baseline approaches such Feedforward Neural Networks (FNN), Bidirectional Long Short-Term Memory (BiLSTM), Gated Recurrent Unit (BiGRU), Gene Expression Programming (GEP). Subsequent feature importance permutation-based strategy, SHAP analysis, revealed humidity hourly volume were influential features. findings indicated medium high values associated an increased likelihood HE, while volumes similarly contributed occurrence HE.

Language: Английский

Citations

0
The Application of ResNet-34 Model Integrating Transfer Learning in the Recognition and Classification of Overseas Chinese Frescoes DOI Open Access
Le Gao, Xin Zhang, Yang Tian

et al.

Published: July 21, 2023

The unique characteristics of frescoes on overseas Chinese buildings can attest to the integration and historical background Western cultures. Reasonable analysis preservation provide sustainable development for culture history. This research adopts image technology based artificial intelligence, proposes a ResNet-34 model method integrating transfer learning. deep learning identify classify source emigrants, effectively deal with problems such as small number fresco images emigrants' buildings, poor quality, difficulty in feature extraction, similar pattern text style. experimental results show that training process proposed this article is stable. On constructed Jiangmen Haikou JHD datasets, final accuracy 98.41%, recall rate 98.53%. above evaluation indicators are superior classic models AlexNet, GoogLeNet, VGGNet. It be seen has strong generalization ability not prone overfitting. cultural connotations regions frescoes.

Language: Английский

Citations

8
Radar Target Classification Using Enhanced Doppler Spectrograms with ResNet34_CA in Ubiquitous Radar DOI Creative Commons
Qiang Song, Shilin Huang, Yue Zhang

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(15), P. 2860 - 2860

Published: Aug. 5, 2024

Ubiquitous Radar has become an essential tool for preventing bird strikes at airports, where accurate target classification is of paramount importance. The working mode Radar, which operates in track-then-identify (TTI) mode, provides both tracking information and Doppler the recognition module. Moreover, main features target’s are concentrated around spectrum. This study innovatively used to generate a feature enhancement layer that can indicate area spectrum located combines it with RGB three-channel spectrogram form RGBA four-channel spectrogram. Compared spectrogram, this method increases accuracy four types targets (ships, birds, flapping flocks) from 93.13% 97.13%, improvement 4%. On basis, integrated coordinate attention (CA) module into building block 34-layer residual network (ResNet34), forming ResNet34_CA. integration enables focus more on target, thereby further improving 97.13% 97.22%.

Language: Английский

Citations

1
A Multi-Scale Content-Structure Feature Extraction Network Applied to Gully Extraction DOI Creative Commons

Feiyang Dong,

Jizhong Jin,

Lei Li

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(19), P. 3562 - 3562

Published: Sept. 25, 2024

Black soil is a precious resource, yet it severely affected by gully erosion, which one of the most serious manifestations land degradation. The determination location and shape gullies crucial for work erosion control. Traditional field measurement methods consume large amount human resources, so great significance to use artificial intelligence techniques automatically extract from satellite remote sensing images. This study obtained distribution map southwestern region Dahe Bay Farm in Inner Mongolia through investigation created dataset. We designed multi-scale content structure feature extraction network analyze images achieve automatic extraction. multi-layer information resnet34 input into module us, respectively, richer intrinsic about image. fusion further fuse structural features improve depth model’s understanding Finally, we muti-scale low-level high-level information, enhance comprehensive model, ability gullies. experimental results show that can effectively avoid interference complex backgrounds Compared with classic semantic segmentation models, DeepLabV3+, PSPNet, UNet, our model achieved best several evaluation metrics, F1 score, recall rate, intersection over union (IoU), an score 0.745, 0.777, IoU 0.586. These proved method highly automated reliable extracting images, simplifies process provides us accurate guide locate gullies, then provide guidance management.

Language: Английский

Citations

1
Classification of Similar Electronic Components by Transfer Learning Methods DOI
Göksu Taş

Published: Jan. 1, 2024

Language: Английский

Citations

0
Automatic Counting and Location of Rice Seedlings in Low Altitude UAV Images Based on Point Supervision DOI Creative Commons

Cheng Li,

Nan Deng,

Shaowei Mi

et al.

Agriculture, Journal Year: 2024, Volume and Issue: 14(12), P. 2169 - 2169

Published: Nov. 28, 2024

The number of rice seedlings and their spatial distribution are the main agronomic components for determining yield. However, above information is manually obtained through visual inspection, which not only labor-intensive time-consuming but also low in accuracy. To address these issues, this paper proposes RS-P2PNet, automatically counts locates point supervision. Specifically, RS-P2PNet first adopts Resnet as its backbone introduces mixed local channel attention (MLCA) each stage. This allows model to pay task-related feature dimensions avoid interference from background. In addition, a multi-scale fusion module (MSFF) proposed by adding different levels features backbone. It combines shallow details high-order semantic seedlings, can improve positioning accuracy model. Finally, two seedling datasets, UERD15 UERD25, with resolutions, constructed verify performance RS-P2PNet. experimental results show that MAE values reach 1.60 2.43 counting task, compared P2PNet, they reduced 30.43% 9.32%, respectively. localization Recall rates 97.50% 96.67%, exceeding those P2PNet 1.55% 1.17%, Therefore, has effectively accomplished seedlings. RMSE on public dataset DRPD 1.7 2.2, respectively, demonstrating good generalization.

Language: Английский

Citations

0
Detecting Mechanical Vibrations in Televisions via Audio Spectrogram Classification DOI

Rômulo Fabrício,

Agemilson Pimentel,

Ruan J.S. Belem

et al.

Published: Nov. 6, 2024

This paper presents a method for contactless detec tion of mechanical vibrations in televisions through audio spec trogram classification, utilizing Convolutional Neural Networks. The model was trained on dataset containing simulated samples and demonstrated high accuracy, with excellent learning curves observed during training. In further evaluation real the performed well, achieving F1-Score rate 99,02% test partition, confirming its potential use preventive maintenance processes addressing issues other audio-dependent equipment, thereby enhancing efficiency quality service.

Language: Английский

Citations

0
Discrepant Semantic Diffusion Boosts Transfer Learning Robustness DOI Open Access
Yajun Gao, Shihao Bai, Xiaowei Zhao

et al.

Electronics, Journal Year: 2023, Volume and Issue: 12(24), P. 5027 - 5027

Published: Dec. 16, 2023

Transfer learning could improve the robustness and generalization of model, reducing potential privacy security risks. It operates by fine-tuning a pre-trained model on downstream datasets. This process not only enhances model’s capacity to acquire generalizable features but also ensures an effective alignment between upstream knowledge domains. can effectively speed up convergence when adapting novel tasks, thereby leading efficient conservation both data computational resources. However, existing methods often neglect discrepant downstream–upstream connections. Instead, they rigidly preserve information without adequate regularization semantic discrepancy. Consequently, this results in weak generalization, issues with collapsed classification, overall inferior performance. The main reason lies connection due mismatched granularity. Therefore, we propose diffusion method for transfer learning, which adjust granularity alleviate classification problem Specifically, proposed framework consists Prior-Guided Diffusion pre-training fine-tuning. Firstly, aims empower semantic-diffusion ability. is achieved through prior, consequently provides more robust classification. Secondly, focuses encouraging diffusion. Its design intends avoid unwanted centralization, causes Furthermore, it constrained discrepancy, serving elevate discrimination capabilities. Extensive experiments eight prevalent datasets confirm that our outperform number state-of-the-art approaches, especially fine-grained or dissimilar (e.g., 3.75% improvement Cars dataset 1.79% SUN under few-shot setting 15% data). sparsity caused protection successfully validate method’s effectiveness field artificial intelligence security.

Language: Английский

Citations

0