Research On Video Text Recognition Of Natural Scene Based On Image Stitching Technology

In the application process of service robots,automatic driving,etc.,video images are often processed to obtain surrounding information.The text contains a lot of high-level semantic information,which plays an important role in the understanding of images and videos.The traditional Optical Character Recognition(OCR)technology is very mature,besides the document,the recognition of natural scenes text is more difficult,and has become a hot field of research.At present,the text acquisition of video is basically divided into single-frame images for processing,which will result in a large number of repetitions and unintuitive,especially the text and digital information of a large field of view,and most methods have very poor results.From this point of view,this article uses the interframe connection to process the entire video,obtain a panoramic view of the text,and obtain intuitive text information.Firstly,the text detection neural network is built,the YOLOv3 target detection framework is modified,the aspect ratio of the anchor frame and the convolution structure are adjusted to make it more in line with the requirements of text detection,and the multiscale anchor frame result is integrated on the dataset such as ICDAR13.An end-to-end training test is conducted to obtain a high-speed and reliable text detection framework.Secondly,a text tracking model is established.For video text processing,the detection of each frame will consume a large amount of computing resources,and tracking instead of detection can improve the video processing speed and obtain key frames.In this paper,the ECO tracking technology is used to continuously track the detected text,proposed improved ECO method,obtain the position change of the text under the motion state,judge the key frame in time,and stipulate the cutting of the detected text and the stop condition of the text tracking.To make the processing of video fully automated.Finally,the key frame images are stitched,focusing on the stitching effect of the text part,eliminating the ghost phenomenon caused by the global stitching,greatly improving the stitching speed and the effect of the text area,obtaining the text panorama and the text information in the video.A local-based global stitching method is proposed,which can obtain the transformation matrix by using the tracked text box area.The whole stitching is associated with the tracking and detection,and the panoramic image can be obtained with the least effective frame.The processing method is consistent for both the large field of view text and the small partial text,ensuring the text information extraction of the entire video processing is intact,which will also solve the problem of repetitiveness of text extraction.The algorithm processing interface is created to facilitate human-computer interaction and better step-by-step processing of video text.