| The detection of text areas in natural scenes is an important application of computer vision,and it also provides some technical support for many applications,such as computer-aided system for the visually handicapped,robot navigation system in urban environment,automatic driving system and so on.Current detection of scanning document has already can satisfy various application scenarios,but for text detection and recognition in natural scene,accounts for only a small part of the image as text,and complicated background of natural scenes,text in natural scene also have all kinds of color,size,shape,the difference between a and light,shade,and the influence of such factors as therefore is a challenging topic.At present,most of the detection methods for natural scene text are based on general target detection method,and some improvements are made to the text characteristics.Through image processing,pattern recognition,deep learning and other technologies,this paper conducts the following research on text detection methods in natural scenes,which is mainly divided into three parts:(1)Text detection method of natural scene based on text contour.Compared with the previous extraction method,MSER method based on multi-channel weak constraint is adopted to extract candidate areas in this paper,which can obtain more candidate areas.In the subsequent recognition work,a method of fusing high confidence text area with low confidence text area is proposed to further filter nontext area,and the character area is naturally grouped into the final required text line area.Quantitative experimental results show that the detection speed of this method is better than other detection methods based on manual features such as stroke width,and its recall rate is higher than other methods because a large number of candidate areas are extracted.(2)Based on per-pixel depth study of the natural scene text detection method,using completely based on the deep learning method to extract natural scene text area,put forward a new full convolution neural network for text detection,based on the center line of the text and methods of the width of the text,to all sorts of text area,contains tilt testing bending text,the text area.Compared with other methods,the detection results are basically better than the previous results on the same data set,and can more accurately describe the text area in the natural scene.(3)The character recognition method adopts the network structure based on the fusion of convolutional neural network and circular neural network.The sequence features of two-dimensional images are extracted through the convolutional neural network,and then the sequence features are input into the circular neural network to obtain the corresponding character probabilities of corresponding positions.Compared with the previous method of character classification based on segmentation and then using neural network,this method reduces many preprocessing and post-processing operations,improves the speed,and improves the accuracy of text recognition in natural scenes. |
