Technical Study Of Text Detection And Recognition In Natural Scene

With the popularization of intelligent devices and development of GPUs,people take more and more photos and begin to find out how to make use of these photos,especially those contain text.Pictures with text can assist people in many applications,such as translation and VR.However,text in these photos are difficult to detect and recognize,because of different shapes,illumination and shadows.In this paper,we introduce the method to detect text in natural scene.Our method can be introduced in two aspects:Firstly,we implement a MSER-based English text detection method,which includes techniques of character candidate filtering,text grouping,and text candidate filtering.We propose an evaluation algorithm of character candidate filtering methods,and introduce an improved technique based on this evaluation algorithm.Moreover,a Deep Convolution Neural Network(CNN)is utilized in text candidate filtering.Experiments demonstrate good performance of our method.Secondly,a MSER-based method is also proposed to detect Chinese text in natural scene.Different from English text detection,we use another filtering technique to filter MSER regions,and also group MSER regions to Chinese word differently.Furthermore,we also implement a CNN-based method to recognize Chinese text.Above all,in this paper,we propose a method to improve English text detection,and also introduce an improved method to detect and recognize Chinese text in natural scene photos.