Text Detection And Recognition In Natural Scenes Based On Deep Learning

The image contains rich information.As the crystallization of human wisdom,the amount of information contained in the text is often much larger than the amount of information carried by the colored texture.Therefore,the detection and recognition in the scene text is very important,but The traditional optical character detection technology is mainly oriented to document images,and the background is simple in text size and font,and cannot be directly applied to natural scenes.Compared with the recognition of English characters,Chinese characters have more types and features,which are more difficult to identify.Based on the shortcomings of current text detection and recognition algorithms,this thesis uses the deep learning algorithm to propose an algorithm for detecting and recognizing Chinese and English in natural scenes.The main work of this thesis is as follows:Firstly,the differences between the classic target detection algorithm and the scene text detection algorithm are analyzed.In view of the shortcomings of the current scene text detection algorithm,this thesis improves a fast and accurate text detection framework(Efficient and Accurate Scene Text Dectecor,EAST):Each scene text produces four kinds of text boundary segments.By semantically identifying the bounding box,you can locate the text pixels around the word or text line instead of all the text pixels.The scene text avoids the positioning of long text.accurate.For the improved detection algorithm,the loss function is redesigned:the confidence of the text pixel,the confidence of the text pixel as the boundary pixel and the distance from the pixel to the text boundary.The local perceptual non-maximum suppression algorithm is improved:the geometric frame predicted by each text box pixel is regressed to form the final text detection frame.Experiments show that the detection algorithm can detect various lines of text in the natural scene.In terms of the average performance(F1-measure),it is 2.56 percentage points higher than the EAST model.Secondly,the natural scene text recognition,to identify the words in the bounding box generated during the detection phase,using the end-to-end convolutional neural network based on image-based sequence identification,has been improved in two aspects:First,the use of deformable Convolution instead of depth separable convolution extraction feature can automatically extract text image features and have better robustness to natural scene text image recognition.Secondly,the post-processing algorithm that converts the final sequence output of the model into a text sequence is improved.The statistical language model is integrated on the basis of the original greedy search algorithm used for decoding.Due to the diversity of Chinese characters,the deep learning method is used.In the process of training it,the existing public data set is far from enough,so this thesis proposes a sample expansion method for natural scene text pictures,which is used to identify the training of the network.Experiments show that the improved recognition algorithm can accurately identify text lines and achieve 89.6%recognition accuracy on the RCTW-17 data set.Finally,this thesis summarizes a series of research work done on this topic,and looks forward to the next research work.