The Research Based On Convolutional Neural Network For Text Detection In Natural Scene Images

At present,natural scene text detection has become an important research issue in the field of computer vision and pattern recognition,document analysis and recognition.However,Natural scene text detection is very different from traditional document image text detection technology.text and background in natural scene images are extremely complex and diverse,which cause strong interference to the segmentation of text and background.Manual design features used in traditional natural scene text detection methods lack robustness in dealing with complex natural scenes.In order to solve the problem of text detection in complex natural things and to extract text effectively,two novel scene text detection methods based on Convolutional Neural Network(CNN)are proposed:multi-directional natural scene text detection method based on Fully Convolutional Networks(FCN)and natural scene text detection method based on multi-channel boundary boxes fusion.For the first method,the Deep Convolutional Neural Network is used as feature extraction network to extract text features in images.Fully Convolutional Networks(FCN)is used to perform up-sampling operations on the basis of multi-layer text feature maps and to merge multi-layer text features layer by layer.Combining with semantic segmentation method,the candidate regions of text are segmented.Then,the text candidate detection boxes are obtained directly from the segmented text candidate region and expanded compensation is conducted.Eventually,the final result is obtained by post-processing the text candidate detection boxes.For the second method,it aims to solve the false alarm problem of the formermethod,therefore,the method of multi-channel acquisition and fusion boundary boxes is proposed.Firstly,feature extraction network is also used to extract the text features in the image and to integrate the feature layer of full convolution network.But when acquiring the text boundary boxes,two independent channel are designed: regression boundaries and direct acquisition of boundary boxes on score maps.The advantages of the two methods are employed to get the final results.The two methods are evaluated in standard datasets such as ICDAR2013 and ICDAR2015.The experimental results show that the two methods have better performance than other latest methods.The multi-directional natural scene text detection method based on FCN perform well on many datasets,but there are some false positives.Compared with the former method,the natural scene text detection method based on multi-channel boundary boxes fusion has improved,solved most false positives,achieved better results on the above datasets and further improved the detection performance.It also indicates that this method is significantly effective in solving natural scene text detection.