Research Of Text Detection In Natural Scene Algorithm Based On Deep Learning

Text as a language information exists in all aspects of our life.According to the captured scene images,the extraction of text information can help us understand the scene,so that it has great social value and significance.However,the complexity of the background and the diversity of the text pose great challenges to the accuracy of the text detection.The traditional machine learning method is gradually processing,resulting in the accumulation of errors and performance bottlenecks.Recently,the development of deep learning brings a great light to the research of text detection.On this background,the paper focuses on the research of scene text detection algorithm based on deep learning.This research includes:(1)The feature information of small-scale text processing through deep network becomes blurred.We design a coarse detector,which combines shallow and deep,local and global information of network,is used to classify each pixel in the original image,and obtain saliency map of text information.It helps to detect small-scale text.(2)The existing method doesn't consider the interrelationship of the detected objects,and the detected text bounding box is redundant.The fine detector proposed is a kind of network by detecting the character or character part.A series of character sequences is detected with a thin bar-shaped anchor adding contextual information.Under the interference of light and so on,it can still detect the complete text.(3)To solve the text detection of the inaccurate positioning,difficulty to detect mixed text on multiple scales and other issues,we design the intermediate-processing mechanism combining the advantages of coarse and fine detectors,and present a cascaded convolutional neural network that locates each line of text more precisely.Through analysis,research and experiment,our algorithm has strong robustness,generalization ability and anti-interference ability in finely positioning squeezed text lines,detecting small-scale texts and detecting partially disturbed texts.