Text Detection From Natural Scene Image

Text detection and recognition has a wide range of applications in the field of computer vision.Text images can be divided into two types:document image and scene image.Compared with the detection of document image,the diversity of the scene image text and the complexity of the background make it difficult to detect the text in scene image.Text detection is the front end of text recognition.Aiming at the problem of text detection in scene images,we study typical text detection algorithms and analyse the advantages and disadvantages of them.Based on the study,we improved MSER algorithm and FASText text detector.The main results are as follows:1.Research on the classical text detection algorithms.Summarize the text detection effects of 21 mainstream algorithms on public data sets ICDAR.The data shows that the existing text detection algorithms still have low recall,and the highest recall rate is only 83%;The advantages and disadvantages of the typical algorithms in the two core steps of candidate text region extraction and text/non-text classification are analyzed.2.MSER text detection method based on Harris corners.To solve the problem of the excessive number of regions detected by the traditional MSER algorithm,a method for traversing all the corner points to extract candidate regions is proposed.Firstly,in order to detect the corner points as completely as possible,the Harris corner detection algorithm is improved.The corner points are detected from the three color channels R,G,B,and all corner points are merged,and the multi-scale space of the pyramid structure is constructed using the nearest neighbor interpolation.Then,according to the principle of the MSER algorithm,an area is generated for each detected Harris corner point.The experimental results show that the improved algorithm can effectively reduce the number of candidate regions on the premise of ensuring no missing text area.3.Improvement of FASText text detector.In order to solve the problem of duplicate detection and non-text area error detection,the pseudo key points are filtered according to the distribution characteristics and color attributes,and use the scan line seed fill algorithm to extract candidate regions.Then decrease text detection by non-maximum suppression.Finally,design a dual threshold classifier for text/non-text classification.The experimental results show that the improved FASText text detector reduces the repetition rate of text detection and improves the effectiveness of text detection.The precision is 91.3%,recall is 72.5%,and F-measure is 81.2%.