Research And Implementation Of Scene Text Recognition Technology Based On Deep Learning

The advent of the era of artificial intelligence has made people’s lives more and more convenient and their happiness index is rising.As the main branch of artificial intelligence,computer vision is developing at an increasingly rapid pace,with research gradually tending in the direction of deep learning.Traditional optical character recognition has been applied to various smart devices,bank card number recognition,ID card recognition,etc.However,the requirements for external factors are extremely high: the content of the captured image must be clear,it is essential to place the ID card avatar in the candidate frame,etc.In addition,optical characters are also used for information entry,reducing human resource consumption and increasing efficiency.Natural scene text is mainly signboards,signage and promotional slogans,characterised by non-fixed lengths,diverse formats and variable sizes.But scene text recognition is still in a period of continuous optimization due to its difficulties.This paper investigates the detection and recognition of text in natural scenes through deep learning techniques,and designs and implements a real-time intelligent detection system for scene text.The main tasks are:(1)To address the problem of insufficient detection of small targets by the YOLOv3 model,the network structure and feature scales are improved:the redundant neurons are removed by the model reduction method of channel reduction to achieve the purpose of model lightweight,and the feature scales of the model are adjusted to improve the detection accuracy of the model for small targets.(2)The 3.6 million datasets are used to train the recognition capability of the CRNN end-to-end recognition model,and the decoding problem of sequences is handled by CTC to recognize non-fixed length text sequences.(3)Based on web programming,the real-time intelligent detection system of scene text is developed to achieve real-time detection and recognition of scene text images.Firstly,the text is detected and its position is determined by the YOLOv3 model,then the text is recognised by the CRNN model,and finally the real-time detection and recognition is achieved and the recognition results are visualised.Experiments on the YOLOv3+CRNN network proposed in this paper are conducted on publicly available datasets and real-life natural scene datasets.The experiments prove that the performance of the end-to-end network proposed in this paper achieves the expected results.Natural scene text detection and recognition technology provides technical support for the development of other branches of artificial intelligence.For example,robotics direction to obtain information through scene text for conversion;intelligent assistance direction,such as blind reading,blind navigation,intelligent driving and other applications.Therefore,it has very important practical significance and value to improve the real-time and accuracy of recognition.