Traffic sign classification with deformable convolution based on denoising residual convolutional autoencoder region localization

Traffic sign classification is particularly important in the field of autonomous driving. Image classification methods based on convolution have been widely studied. Traditional convolution, due to its fixed sampling and operation mode, is difficult to capture information in specific irregular key a...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Engineering Research Express Ročník 7; číslo 4; s. 45257 - 45276
Hlavní autori: Pan, Hao, Guo, Qianlu, Yuan, Decheng, Pan, Duotao, Li, Dong, Yu, Qianxin
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: IOP Publishing 31.12.2025
Predmet:
deformable convolution
image region localization
multi-head self-attention mechanism
residual convolutional autoencoder
traffic sign classification
ISSN:2631-8695, 2631-8695
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Traffic sign classification is particularly important in the field of autonomous driving. Image classification methods based on convolution have been widely studied. Traditional convolution, due to its fixed sampling and operation mode, is difficult to capture information in specific irregular key areas. Although deformable convolution solves the fixed sampling problem of standard convolution, the images captured in reality are easily disturbed by noise, leading to incorrect offset learning and thus destroying the spatial consistency of features. To address these issues, this paper proposes an image classification method based on denoising residual convolutional autoencoder for important region localization and deformable convolution (DRCAE-DCN). The denoising residual convolutional autoencoder is used to extract feature information from the images with added noise and restore them to clean images without noise. The importance of the extracted features is calculated through a multi-head attention mechanism, and the features with high importance are selected to locate the regions in the original image. Deformable convolution is then applied to these regions. By screening important features and localizing regions in the image, it provides a basis for the offset learning of deformable convolution. The important features are used as auxiliary information to fuse with the features extracted by deformable convolution, compensating for the deficiency of deformable convolution in global information extraction, and then image classification is performed. The classification accuracy on the GTSRB, CCTSDB and BelgiumTS datasets has been significantly improved, demonstrating the effectiveness of this method.
Bibliografia:ERX-110424.R3
ISSN:2631-8695
2631-8695
DOI:10.1088/2631-8695/ae1886