Research On Character Detection Method Of Natural Scene Based On YOLOv3

Character detection plays an important role in many aspects of life,such as identity authentication,translation and vehicle license plate recognition.Natural scene images are diverse.The huge differences in font size,font color and background make the scene character detection a challenging problem.In the early years,character detection relied on manual extraction of character features in images.This method can only be used in simple background scenes such as document recognition,and it is also a time-consuming and laborious work to define and extract features manually.The development of traditional character detection technology has encountered a bottleneck.With the development of deep learning technology,researchers began to use convolutional neural network to automatically extract image features,character detection technology has a new breakthrough direction.In recent years,many researchers regard character detection as a special case of object detection,regard character in image as object,and use various object detection algorithms based on deep learning to detect characters in natural scenes.The technology of character recognition in natural scene can be divided into two independent sub problems: location and classification.Location solves the problem of character location and classification solves the problem of character category.There are two problems in character recognition model of natural scene based on deep learning technology.One is how to locate the character region quickly.The other is how to improve the detection accuracy.In order to solve these two problems,this paper proposes a new character recognition model for natural scenes.The main innovations of this paper are as follows.(1)This paper proposes a new character detection model for natural scenes.Considering the character in the natural scene,the character object takes up a small part of the whole image compared with other objects.Therefore this paper improves the YOLOv3 network and designs a new network for character recognition called TYOLO(Tiny YOLO)to detect small targets.The network can not only accelerate the speed of detection,but also improve the accuracy of character detection.In addition,in the previous network model of object detection,only a few types of anchor boxes are manually set,and the manually set anchor boxes may not exactly meet the object size.In comparison,this paper uses clustering algorithm to calculate the geometric characteristics of characters in database,and automatically select anchor box according to the statistical results,which can more accurately locate the character position.Experimental results on an open public database show that the proposed model is superior to other state-of-art algorithms.(2)This paper proposes a new loss function for character detection in natural scenes.Generally,the number of characters in the natural scene character image is only a few.In the stage of extracting anchor boxes,only a few anchor boxes with characters will be generated,which leads to the imbalance of positive and negative samples.To solve this problem,we propose a new loss function,which can adaptively reduce the weight of negative samples in training.In addition,this study no longer uses the traditional intersection over union to calculate the location loss,but proposes a new location loss function,which fully considers the situation that the prediction boxes and the true boxes do not intersect,which is conducive to improving the location accuracy of the model.From the point of view of how to increase the diversity of training data,this study also proposes a new data augmentations method to create a variety of samples and enhance the generalization ability of the model.(3)This paper presents a new post-processing method.In the process of character detection,some special characters are easy to be predicted into multiple objects.For example,the left half of character m is easy to be detected as n,which results in the low detection accuracy of the whole algorithm.The post-processing method proposed in this study considers the above situation and solves the problem that single object is predicted to be multi objects.