Research on Text Detection Algorithm based on Improved FPN

Abstract

In recent years, due to the increasingly application of machine vision in various aspects, the topic of text recognition in actual scenes has gradually become a research hot spot of machine vision research. For images with complex backgrounds, the first thing to do is to accurately locate the position of the target text, and then the text content can be efficiently identified. However, as far as the current text detection algorithms based on deep learning are concerned, there are still problems such as incomplete extraction of text feature regions, and wrong detection of images as text regions. Therefore, this paper proposes an improved DB algorithm to solve the problems of the variable shape and complex background of the text area in the label text detection task, so that the algorithm can achieve better detection effect and better performance in complex scenes. The content of the article is mainly from the following main aspects:firstly, the current situation of text detection algorithms is introduced, then the improvement of the DB algorithm with ResNet-50 as the backbone network is proposed, and finally mPA (mean Average Precision) is used as the evaluation of text detection. By comparing the various detection algorithms, it is found that the improved algorithm has significantly improved the detection accuracy and recall rate, and the model speed is also faster.