A Research Of Text Detection In Natural Scene Images

With the rapid growth of network image analysis and the wide application of intelligent devices,the research of image processing method for acquiring text information in natural scene is becoming more and more active.As the first step in the acquisition of text information in natural scenes,the text detection algorithm in natural scene is the key component.This thesis focuses on the text detection algorithm in natural scene,mainly studies the synthesis of natural scene text picture,the text detection algorithm based on convolution neural network,and the specific contents are as follows:1.We propose an improved algorithm for synthesizing natural scene text images,which can quickly generate a large number of natural scene text images with accurate annotation information and effectively solve the problem of scarce training samples in the process of realizing natural scene text detection.The continuous region segmentation information and depth information of images obtained by using gPb-UCM algorithm and FCRN-DepthPrediction algorithm respectively can significantly improve the effect of text fusion between images.2.We study a text detection algorithm in natural scene based on deep convolution neural network,through reconstructing the ResNet network by using feature pyramid structure and RoIalign region pooling,the feature extraction of small target is optimized,and the RPN(Region Proposal Network)is improved to make it more suitable for the detection of text target.The combination of the two algorithms is applied to the realization of Mask R-CNN target detection algorithm,which effectively improves the detection effect of text target.3.We research on the effects of variables such as data,hyperparameters and training strategy on model training experimentally,and the performance difference between this algorithm and other algorithms is compared and analyzed on the platform of ICDAR2013 competition.To sum up,by using the improved synthetic natural scene text picture and the improved implementation of Mask R-CNN algorithm,the proposed algorithm is compared with the existing algorithm in the general data set,and obtains good results.