Global Feature Fusion and Double-layer Lightweight Crack Sample Augmentation Algorithm
XIE Yonghua
LI Jianyuan
CHEN Ya
ZHUO Annan
Abstract：As a common problem of small sample identification,tunnel crack detection based on deep learning will cause low classification accuracy due to too few crack samples.Based on the CycleGAN model,a crack sample augmentation method GDCycle-GAN based on the global feature fusion double-layer lightweight model is proposed.The GFF(Global Featue Fusion)module is in-troduced into the generator of CycleGAN,which integrates the semantic information of the upper and lower layers of the crack image to obtain multi-scale information,and uses the attention mechanism to strengthen the selection of the crack image feature informa-tion to weaken the crack background information.Aiming at the problem that the efficiency of sample generation is not high,the DL(Double Lightweight)module is introduced into the discriminator of CycleGAN,and group convolution and deep separable convolu-tion are used instead of the original convolutional layer to reduce the amount of convolution operation parameters of each layer and improve the quality of generated samples and the speed of network training.Experimental results show that after the small sample ex-pansion of the proposed algorithm,the speed of expanded sample training and the subsequent crack classification accuracy are im-proved.
Keywords：deep learningcrack detectionfeature fusionattention mechanismGDCycleGAN model
Publication Date：2025-12-20
Online Publishing Date：2026-03-09(First online date of this platform, not the publication date of the document)
Pages：7( 3299-3304,3312 )