Research On Natural Scene Text Detection Based On SSD Algorithm

The field of computer vision has always attracted researchers to explore continuously.The technology in this field can be applied to production control,smart city,information extraction and many other aspects.The detection of text in natural scenes is one of the key technology to extract high-level text information from images.In recent years,along with the development of deep learning technology,many excellent natural scene text detection methods have been proposed,but how to combine the advantages of these methods to achieve better detection methods,there is no simple and unified solution.Based on the study of various natural scene text detection models,this paper proposes a detection model based on SSD algorithm.The model uses the proposed External Mutual Correction method to combine two classical text detection frameworks: semantic segmentation and object detection.The proposed fusion method simultaneously runs the semantic segmentation algorithm and the object detection algorithm to obtain the text detection results respectively,and then their results are used to do mutual correction processing.In order to improve the existing text detection method,this paper first improves the SSD algorithm by introducing a Hierarchical Inception structure,and the size of the default text bounding boxes are modified according to the characteristics of the natural scene text.After obtaining the detection results of the improved SSD algorithm and the semantic segmentation algorithm,the proposed Bounding box Enhancement Module is used to calculate the regional median probability of the SSD detection results by using the semantic segmentation results,and the text bounding box is discarded if the probability is low.In order to further use the semantic segmentation results,this paper proposes a Semantic Bounding box Module,which uses the dense condition random field to process the textual adhesion and misjudgment in the semantic segmentation results,and finds the text bounding box coordinates.Finally,the SSD algorithm and the semantic segmentation detection result are merged again by using the non-maximum suppression,and the advantages of the two methods are fully combined to obtain the optimal natural scene text detection results.In order to evaluate the performance of the natural scene text detection based on SSD algorithm,the standard dataset ICDAR2013 and Street View Text are used to train and test the proposed algorithm.The experimental results show that the external mutual correction method proposed in this paper is superior to the traditional fusion method.The external mutual correction method increases the F-measure score by 13.26%,and the fully improved method increases the F-measure score by 22.01%.The highest F-measure score is 83.68%.It is clear that the proposed method can combine the advantages of the two methods to obtain better natural scene text detection results.