| Writing,as a recording symbol of human information,can express human thoughts and describe things completely through the limitations of time and space.There is a lot of text information in the natural scene,which plays a positive role in the description and understanding of the scene content,and helps people to better understand the content of the scene.Natural scene text detection is a technology that locates text area intelligently in natural scene environment.At present,the common methods of natural scene text detection are mainly based on regional regression,semantic segmentation and mixed methods based on regional regression and semantic segmentation.On the basis of combing natural scene text detection methods and related detection models,thesis conducts in-depth research on how to improve the accuracy rate of detection by improving the network model and how to reduce the number of labeled samples.The main research work is summarized as follows:(1)Adaptive feature enhancement network is proposedAt present,in the field of natural scene text detection,the size of text varies greatly,and there are cases of missing detection in small text areas and under-detection in large text areas.To solve these problems,a multi-core pooling module and an adaptive feature enhancement module are proposed.The feature pyramid network is used to capture the feature information of small targets,and the multi-core pooling module is used to capture more context information in the bottleneck area of the model,which alleviates the information loss in the downward propagation of high-level semantic features and enables the network to obtain more semantic information.In ICDAR2015 data set,compared with DBNet model,the accuracy rate,recall rate and composite index increase by 0.12%,0.48% and0.27% respectively,which proves that the adaptive feature enhancement module can supplement and enhance the information in the fusion feature map at different scales,and reduce the conflicts generated during the fusion of text feature information at different levels.(2)Thesis proposes the integration of multi-scale spatial coordinate attention moduleNatural scene text detection is affected by noise such as illumination and complex background,which brings challenges to the task of scene text detection.Based on the idea of Attention mechanism,an Integrated Multi-scale Spatial Coordinate Attention Module(MSCA)was proposed.In ICDAR2015 data set,compared with DBNet model,the accuracy rate,recall rate and composite index increase by 096%,0.64% and 0.75%,respectively.Experimental results show that the introduction of MSCA module into Res Net+FPN structure can learn the importance of different features after fusion.Thus,the expression ability of feature information and the discrimination ability of text area can be effectively improved,and the false detection of non-text area can be reduced.(3)Text detection network based on semi-supervisionIn the scene text detection task based on semantic segmentation,there are usually few labeled data sets and the cost of manual annotation is high.Based on this,thesis explores a semi-supervised model to solve the problem that there are few labeled data sets and the network model cannot be fully trained.Based on the idea of generating adversarial network,thesis uses the discriminator to generate pseudo-tags in generating adversarial network to help generate network for supplementary training,and explores the influence of different labeling rates of data set tags on segmentation results.In ICDAR2015 data set,compared with DBNet model,the accuracy rate,recall rate and composite index increase by 0.57%,0.87% and 0.69%,respectively.Experiments show that under the condition of full supervision,it is verified that the adversarial training can improve the performance of segmentation network. |
