Research On Weakly Supervised Scene Text Detection Method Based On Text Centerline And Character Heat Map

As an important tool for communication and collaboration,text plays an important role in society.At the same time,in modern society,due to the rapid development of digital content such as pictures and videos,text detection and recognition can often provide people with very important information,and the use of this information can help the subject who uses the information to better complete the task.Therefore,it is of great significance to detect and recognize scene text in pictures and videos.In the process of scene text detection and recognition,it is the primary task to locate the text in the scene(i.e.scene text detection),so scene text detection has important research significance.The content of this research is to detect the scene text in the picture.In this paper,using the center line of the text in the process of study and character heat map as text forms,and use the encoding and decoding network as the backbone network,which in the encoding and decoding network,due to the different decoding output layer with different scales,the different scale of output has a different activation on category information.Therefore,in order to make better use of the category information of different decoding layers,this paper carries out local feature fusion for the features of adjacent decoding layers in the decoding structure,and uses the network structure based on self-attention mechanism to enhance the features after local feature fusion.The experimental results show that the scene text detection effect using the text center line and character heat map as the text representation,and combining the features of the adjacent two layers in the decoding structure for local feature fusion and feature enhancement can be improved compared with the original scene text detection effect.In order to make better use of the semantic information contained in the encoding and decoding structure,only the fusion enhancement features of adjacent decoding layers are included in the features of local feature fusion enhancement.In this paper,global feature fusion is adopted to further fuse the last layer features in the coding structure with those enhanced by local feature fusion,and the fused features are enhanced by self-attention mechanism based network.The experimental results show that after the local feature fusion and enhancement,the global feature fusion and enhancement continue,and the effect of scene text detection will be further improved.In this paper,segmentation prediction method is adopted to realize scene text detection instead of directly regressing text box,so the text instance generation method is needed to convert the corresponding text segmentation results into corresponding text boxes for subsequent test evaluation and achievement display.Considering the text representation method adopted in this paper and the problems existing in the existing methods,a text instance generation method suitable for text center line and character heat map is proposed to form text border.The experimental results show that compared with the original text instance generation method,the detection effect of the proposed method will be improved to a certain extent.