Lightweight Neural Network Based Text Detection for Terminal Blocks
Aiming at the problem that the current deep learning methods have low accuracy and a huge model for terminal row detection,which is difficult to deploy on the mobile side and thus cannot realize industrialized application,a lightweight neural network-based text detection model for terminal row is proposed.Firstly,PSENet improves the backbone network and proposes a backbone network that extracts fused image features and edge features to improve the expression ability of the features;secondly,it proposes a mechanism that combines spatial attention and channel attention embedded in positional information to enhance the ability to extract important information;lastly,it improves the feature pyramid network in a lightweight way and reduces the number of parameters.Tests on the terminal row labeling dataset yield results:compared with the PSENet model,the improved model increases accuracy by 3%,the F1-score by 1.5%,and the model size is reduced by 37.72%.The experimental comparison shows that the model significantly improves the accuracy rate and lightweight level compared to other mainstream models,and the results of the study can provide technical support for detecting terminal row labels on mobile.
text detectionattention mechanismdepth separable convolutionMobileNetedge detection algorithm
万方数据
维普
