Research On Fusion Techinques Of Deep Learning Based Scene Text Detection Algorithms

With the popularity of various digital video cameras and mobile phone cameras,a large amount of image and video data are generated every day.These massive image and video resources provide a large amount of training data for researchers in the target detection direction scene text detection,which is an important direction in computer vision.People can use scene text detection algorithms to extract images and video information and apply these techniques to people's daily life.The research object of this paper is the integration of deep learning based natural scene text detection algorithms.It can be applied to the bill detection and ticket identification tasks in the banking system.It can also be utilized by network police to automatically identify sensitive vocabulary and maintain network security.The text in screenshots of the fault in the banking system can be detected by the ticket through the text detection algorithms,Which is helpful in improving the text recognition rate and the search ability of the case library.However,the images in the natural scene usually have Chailenges such as uneven illumination,occlusion of text,inconsistent text size,and the large number of fonts.These problems increase the difficulty of scene text detection.After research of more than a decade,people have proposed a variety of natural scene text detection algorithms,but constrainted to specific data sets used,many algorithms do not have a good generalization ability on different scene text datasets.In this paper,the text detection task in the natural scene is studied.The aims of this work include the following three aspects.First,we collect and annotate the Chinese scene image data set(ShopSign).Second,a feature fusion based scene text detection algorithm is designed.Finally,a method that integrates the detection results of various scene text detection algorithms is designed.First of all,we collect and label a large-scal Chinese scene image data set(ShopSign).Scene text dataset is the most basic part of scene text detection technology.Although there are many text detection datasets at present,most of them are English datasets,and few of them are in Chinese,so we capture,collect — a large dataset using smart phone.Finally,there are 25,770 images in ShopSign,of which 96% were taken by hands.After a lot of manual labeling work,these Chinese scene image images are annotated and multi-calibrated to form the Chinese scene image data set: ShopSign.Secondly,we design a feature fusion based scene text detection algorithm.The feature extraction networks VGG16 and Inception are integrated into the EAST(An Efficient and Accurate Scene Text Detector)algorithm to form a scene text detection algorithm with multiple feature extraction networks.The experiments are designed and related experiments are carried out.Third,a scene text detection algorithm that combines heterogeneous method detection results is proposed.In the design of the algorithm for the fusion of heterogeneous method detection results,this paper analyzes the advantages and disadvantages of the current text detection algorithms in scene text detection.After comprehensive comparative analysis,this paper selects EAST,TextBoxes++,CRPN(Corner-based Region Proposals Network)and RRPN(Rotation Proposals Network)algorithms as the baseline algorithms.When the detection results of the four algorithms are combined,this paper considers that there are at least two detection frames in the same area of the same picture,and it is considered that there is a text line in the area.After overlapping,the result of one detection frame of the area is retained.In addition to using the ShopSign dataset,this paper also uses the English horizontal dataset ICDAR2013,the English multidirectional dataset ICDAR2015,the MSRA-TD500,and the Chinese multi-directional dataset RCTW as the datasets of the experiment.In summary,this paper collects and organizes the Chinese scene image dataset ShopSign,which provides a data foundation for the development of natural scene text detection direction.The scene text detection algorithm based on feature fusion is designed,and good detection results can be obtained.At the same time,this paper designs a scene text detection algorithm that combines the detection results of heterogeneous methods.By analyzing and comparing the detection results of the four text detection algorithms and the fusion detection results,it is finally found that the fusion methods can achieve better performance.This research results provides a alternative idea for the natural scene text detection research,and has certain reference value for the development and applications of scene text detection.