Research On Cascaded Text Detection In Natural Scene Images Based On SSD

Image is an indispensable information carrier in human daily life and work,which is the important bridge for the information to interchange,according to statistics,there is 75% information is visual information that gains by image for everyone.The text in image is the important cure to understand and describe the scene content,what's more,an accurate and efficient nature scene text detection and recognition can provide important assistance in a lot of fields,such as intelligent image retrieval,robot navigation and unmanned vehicle,so the nature scene text detection and recognition has become one of the key points of development in many fields.As the cornerstone and premise of text detection and recognition,the main task of Text detection is to locate the text in the image and it has become one of the most important issue in the fields of computer visual.In recent years,with the development of deep leaning and reduction of computer hardware cost,CNN becomes the breach of text detection,and object detection algorithm based on deep learning has become one of the research directions of text detection.But because of the difference of the text area in the image,object detection algorithm doesn't make well when detecting text in the nature scene,as it caused two problem that is missed detection on little text area and incomplete detection on long and narrow text area in the image.So,the paper improves SSD and uses the idea of cascaded CNN to design cascaded text detection in natural scene images based on SSD.First,the paper analyses the distribution characteristics of the text and the model of the SSD,so that explains the advantages and problems of SSD when it detects text area in natural scene.Then,the paper design a model of text detection,which has better robustness when the text is little or long and narrow.Finally,the paper uses the idea of cascaded to classify the candidate text region by a model of text classification,and gains the final result by remove the non-text region.The experimental results show that the text detection proposed in this paper make well when detecting natural scene image with complex background,and especially for the problem of missed detection on little text area and incomplete detection on long and narrow text area.And the experimental results in ICDAR 2013 shows that the recall is 78.6%,the precision is 85.1% and the F-mean is 81.7%,which proves that our method can effectively detect text from the natural scene image with complex background.