A Research On Irregular-Shaped Text Detection Algorithm In Natural Scene Images

Scene text detection refers to locating the text area in a scene image or video and marking it with a bounding box.One of the most challenging tasks is to detect arbitrary-shaped text.It requires a polygonal bounding box that closely surrounds the text area.This kind of text representation is more complicated and can meet the needs of reality.At present,scene text detection algorithms of arbitrary shapes can be divided into top-down algorithm based on instance segmentation and bottom-up algorithm based on semantic segmentation.However,there are still some problems in current work.For example,the post-processing process of the algorithm based on semantic segmentation is complex;conventional text detection algorithms are quite sensitive to threshold selection;conventional text detection algorithms rarely consider unobvious boundaries of text objects;for some text detection algorithms,when the input resolution becomes larger,the text line will be detected as several broken words.In view of the above problems,this paper researches the algorithm of detecting arbitrary-shaped text in natural scene images from three aspects: model fusion,text edge processing and multi-scale feature map fusion.The main research contents and contributions of this paper are as follows:1.This paper proposes a text detection algorithm that combines multiple representations.To improve the complicated post-processing process of the semantic-segmentation based algorithm,this paper introduces a text kernel enhanced algorithm that combines the output of semantic segmentation branch and instance segmentation branch.Experiments show that the algorithm can effectively reduce the pixels to be processed in the post-processing stage by more than 35%,thereby reducing the time overhead of post-processing.2.This paper proposes a text detection algorithm that is insensitive to threshold selection.Conventional detection algorithms are sensitive to threshold selection.In order to solve this problem,this paper introduces an edge weight and size weight that gradually decay with the training process when calculating the loss function,which effectively reduces the adverse effects of uncertain labeling of text edges.Due to the accurate prediction of the text kernel area,the algorithm can eliminate the complicated post-processing process,and run faster than most existing algorithms.Experiments show that the algorithm achieves an F1-score of 86.1% on the data set Total-Text and an 89.3% F1-score on the data set MSRA-TD500.At the same time,the algorithm is insensitive to threshold selection,and the performance is stable under the selection of different thresholds.3.This paper proposes an adaptive fusion multi-scale feature map text detection algorithm.Aiming at the problem that conventional algorithms cannot detect complete text lines in high-resolution images,an adaptive fusion module is introduced in this paper to fuse two feature maps that are up-sampled to different scales into one feature map.The model can adaptively adjust the weights of feature maps with different resolutions by learning the context of the environment.Experiments show that,compared with the baseline model,when the input image resolution is higher,the accuracy of the results is steadily improved by more than 4% with almost no loss in detection efficiency.