| Reading text from natural scene images has received extensive attention in the fields of computer vision,pattern recognition,and industry due to its wide application.In the era of intelligent big data,the ability to effectively extract and analyze image text content and make information intelligent has profound significance for theoretical research in academia and technology-driven in industry,and image text detection is the prerequisite for subsequent text content recognition Conditions,but due to many problems such as illumination,shooting angle,background and text diversity,currently there are few technical methods that can be applied to the text detection of complex scene images.Therefore,to improve the robustness and accuracy of localization and detection in complex scene text has positive research significance and important application value.The content and contributions of this topic are as follows:First,considering that the different levels of features in the image have different highlighting effects on the target’s category and location information,a network architecture that combines the multi-scale features of the image with Residual Network(Res Net)and feature image pyramid(FPN)is proposed.The low-level features with medium and high resolution but weak semantic expression ability and abstraction are higher but the high-level features with stronger semantic expression are extracted at the same time,and the multi-scale information is fused using the special structure of FPN three-segment connection to generate Multi-level feature map;Secondly,in order to match the proportion segmentation masks of different levels of feature maps,a text real value label generation algorithm based on polygon reduction algorithm is proposed to generate segment regions of different proportions.In the loss function,hyperparameters are used to balance the impact of the original size text segmentation instance and the reduced text segmentation instance,and two loss functions are designed,one is based on the binary cross entropy loss function(BCE),and the other is based on Online hard case mining(OHEM)and dice coefficient loss functions,and compare the effects of the two loss functions on the network model through experiments;Finally,in the post-processing step,in order to expand the text instance d from segmentation masks of different scales,a progressive size expansion algorithm based on watershed is proposed,using the "waterlogging" feature of the watershed algorithm to sequentially access larger text segmentation Version,iteratively annotated to perform pixel-wise connected region expansion on the smallest segmentation instance until the largest segmentation result is obtained.The related technical scheme of scene image text detection based on multi-scale feature fusion and instance segmentation proposed in this study has been experimentally verified to show good robustness for the detection of directional text,multilingual text and curved text.The F-Score on ICDAR2015,ICDAR2017-MLT,CTW-1500,Total-Text data sets were 82.32%,70.88%,79.1%,78.9% respectively. |
