DocPointNet: 3D text line detection method for point cloud
Fuping Su · 2024
With the advancement of the digitalisation wave, the importance of 3D point cloud data is becoming more and more prominent. In response to the current lack of 3D point cloud text line datasets, this paper constructs the DocPointCloud dataset, which is generated by 3D reconstruction of document images and provides rich data resources for 3D text line detection. At the same time, for the problem that existing approaches cannot accurately detect 3D text lines when processing document point clouds, this paper proposes the DocPointNet model. By introducing the feature pre-extraction module and the attention mechanism, this model significantly enhances the feature representation ability of the model on point cloud data, thus achieving the accurate detection of text lines in point cloud. Experimental results show that the mIoU of the model on the DocPointCloud dataset is 0.72, which verifies its effectiveness in the 3D point cloud text line detection task.