| Traditional optical character recognition methods are relatively mature and have been applied to document scanning and other fields.Natural scene text images usually contain complex backgrounds,various fonts,uncontrollable illumination,and factors that are not conducive to traditional text detection and recognition,such as perspective transformation and distortion,which lead to a significant performance degradation of traditional methods in processing natural scene text,and the detection and recognition of natural scene text has some room for improvement.This dissertation focus on the research and analysis of the problems in scene text detection and recognition based on deep learning,and the specific work is as follows.To achieve accurate detection of scene text of arbitrary shapes,a network for complex scene text detection based on instance segmentation method is proposed.The features are extracted using Mi T,a backbone network specially designed for instance segmentation,and the feature pyramid network in the feature fusion part is improved to perform feature fusion and obtain pixel-level segmentation results by using a multilayer perceptron.The problem of poor detection of curved text and text with arbitrary rotation angle in scene text detection is mainly solved.To extract richer text semantic information to improve the performance of scene text recognition,a scene text recognition method that fuses semantic information is proposed.To better extract image features,a feature extraction network Red Net47 is designed using a combination of Involution and convolution and decoding is performed using Transformer.The language model is improved with Reformer to make it more lightweight,and the improved language model is used to correct the scene text detection results.The problem that individual character recognition errors caused by illumination and blurring have an impact on the recognition results of the whole word is mainly solved.To solve the problem that scene text detection and recognition need to be performed independently,an end-to-end scene text detection and recognition network is proposed.The text detection part is improved with a lightweight feature extraction network and a Transformer-based pyramidal feature fusion network.The text recognition part uses Masked Ro I to connect the detection and recognition parts,and makes the language model in the recognition part iteratively optimize the recognition results.The proposed network enables end-to-end detection and recognition. |
