Research On Text Object Detection And Recognition In Natural Street Scene

The way people in a natural street scene obtain information is their five natural senses and their own perceptions under innate conditions.Among them,vision is especially critical for the information processing in natural street scenes,and among the many information that can be seen in sight,the most intuitive and easy to understand is textual information.Advances in science and technology have made text object detection and recognition technology more and more developed.This paper takes natural street scenes as the research background and conducts text target detection and recognition research in complex scenes.The segmented traditional text detection and recognition method is analyzed.Aiming at the shortcomings of the existing technology for small text target detection and recognition,a small text target optimization method under natural street scene is proposed.Aiming at the problems of small text annotation data,large manual labeling workload and low efficiency,a semi-supervised learning algorithm is proposed.The main research work is as follows:1)Research on the optimization method of small text targets in natural street scenes.Construct a three-level training data set for small text target detection and recognition,and design an intensive training model from easy to difficult.A text target image DCT coefficient synthesis method is proposed to synthesize the first two training data sets.Aiming at the shortcomings of the existing technology when processing small target texts,a small text target optimization method based on resolution compensation is proposed,and the effectiveness of the optimization method is verified through experiments.2)Propose a semi-supervised text target detection and recognition algorithm.The algorithm adopts a parallel structure of text target detection network and recognition network,and through preprocessing and feature extraction link to ensure that text target detection and recognition can share features:so that the result of text target recognition is not affected by the text target detection frame.In this paper,a text target detection and recognition method based on Bayesian network is proposed to perform feature extraction,label prediction and screening for samples.By comparing and evaluating the performance of existing algorithms on multiple public test sets,the effectiveness of this algorithm and the superiority of dealing with small text targets are verified.