Scene Text Detection Algorithm Based on Non-local Feature Enhancement
To precisely distinguish adjacent text instances in natural scenes,quickly and accurately locate text instances,a text detection algorithm with non-local feature enhancement is proposed.Based on DBNet,the algorithm takes the lightweight net-work resnet-18 as the backbone network,and adopts feature pyramid enhancement module(FPEM)and feature pyramid fusion module(FFM)to compensate the insufficiency of feature extraction capabilities of lightweight networks.Considering the limitation of traditional network processing area,an improved non global context net(GCNet)is introduced,so as to facilitate the model to capture image information from a global perspective.Meanwhile,a differentiable binary optimization segmentation network is adopt-ed to distinguish close text instances,and the post-processing is simultaneously simplified.For the unbalanced data samples,the fo-cal loss is selected as the loss function,the weights of positive and negative samples are adjusted,and pay more attention to difficult samples.The experiment results show that the algorithm achieves a certain improvement on value F and detection speed on IC-DAR2015 data set if comparing with current advanced DBNet.
text detectionfeature enhancementGC Netdifferentiable binarizationfocal loss
国家科技期刊平台
NSTL
万方数据
