Scene Text Detection Based On Instance Segmentation

As an important way of information dissemination in the history of human beings,text directly carries rich and advanced semantic information.Text detection in natural scene images is of great help for visual tasks such as scene understanding,image retrieval,human-computer interaction and so on.Although methods of text detection in electronic document images has been mature,word detection task still faces severe challenges in more complex scenarios.In this thesis,we combine deep learning techniques to explore the task of text detection in natural scene images.In order to make the system robust in complex scenes,a text detection algorithm based on instance segmentation is proposed.We consider our target of text detection to be a patch with a single Latin word as a unit.In our algorithm,the pipeline can be divided into three steps:text feature maps prediction,text instance segmentation,and word bounding box calculation.Among them,text feature maps include text saliency map and text instance windmill map,which are jointly predicted by a multi-task deep fully convolutional network.On the basis of text feature maps,we generate text instance segmentation with a designed graph pruning method.In order to compare with other algorithms,we calculate text instance bounding boxes as a post-process.In practise,we implement a scene text detection system based on the proposed method,which locates the position of each text instance and generates corresponding bounding box for each instance.This system can be applied to images with arbitrary shooting conditions,background scenes and text forms.A series of experiments on standard benchmarks,mostly ICDAR 2015 are conducted on the proposed text detection algorithm.Our evaluation for the method includes accu-racy,recall and harmonious average.The experiment result demonstrates that instance segmentation based text detection algorithm outperforms the existing algorithm significant-ly,which shows the effectiveness and superiority of the proposed method.In addition,we compare instance segmentation methods with general object detection methods,and analyse the advantages of instance segmentation in text detection scenarios.