Text Detection In Natural Scenes Based On Improved TextBoxes

Text detection in natural scenes has important research value and application significance in computer vision and image processing.Due to the complex background in natural scenes,traditional text detection algorithms require a large number of heuristic rules to filter out text regions and the effect is not obvious.At present,TextBoxes algorithm based on deep learning can directly predict text with different scales,that is the low-level feature layer of TextBoxes network predicts small-scale texts and the high-level feature layer predicts large-scale texts.Moreover,a nonmaximum suppression algorithm(NMS)is used to process the candidate text boxes in order to acquire the final text detection result.However,the low-level feature layer of the TextBoxes network has weak presentation and the accuracy of predicting smallscale texts is relatively low in the TextBoxes algorithm.In addition,the non-maximum suppression algorithm gains unsatisfactory results in processing the candidate text boxes.To deal with the deficiency of TextBoxes algorithm,an improved TextBoxes text detection algorithm is put forward,including an improved TextBoxes network and an improved non-maximum suppression algorithm.The main work of this paper includes:1)For the problem that feature layers of TextBoxes network have weak presentation, an improved TextBoxes network is proposed.A new neural network is built by adding deconvolution layer and Eltwise layer inside the original network and fusing different feature layers of original network.The improved network improves the performance of text detection.2)For the problem that small-scale texts detection is not robust enough in Text Boxes network,high-level features are integrated into the lower layers of neural network and three network structures are presented.This three network structures are two adjacent feature layers merged,three adjacent feature layers merged,and the highest feature layer is merged with other feature layers respectively.Among them, the network structure where the top feature layer merges with other feature layers greatly improves the performance for small-scale text detection.3)For the problems that redundant text boxes can not be removed effectively and the detection results are incomplete in non-maximum suppression algorithm,an improved non-maximum suppression algorithm is proposed,called Text Bounding Box Fusion algorithm(Text-BBF).This algorithm uses an including overlap rate to remove redundant text boxes,and makes the most of the position information of neighboring candidate boxes to fuse candidate boxes.The improved algorithm can obtain more precise text position and improve the performance of text detection.In this paper,the improved TextBoxes algorithms are validated on ICDAR2011 and ICDAR2013 datasets.There are there improved algorithms: the algorithm that improved TextBoxes network combined with NMS,the algorithm that TextBoxes network combined with Text-BBF,the algorithm that improved TextBoxes network combined with Text-BBF.The experimental results show that the improved Text Boxes network can effectively detect small-scale text,and the Text-BBF can further improve the accuracy and recall rate of text detection.This three algorithms presented in this paper have higher positioning accuracy and robustness in the natural scenes text detection.Among them,the text detection algorithm that improved Text Boxes network combined with Text-BBF is superior and the F-Measure has increased by 4% on two dataset compared with the TextBoxes algorithm.