A Study on Tibetan Detection in High-Resolution Thangka Images Utilizing SAC-DBNet

In response to the challenges posed by high- resolution images, including significant text scale disparities, intricate image backgrounds, and a low proportion of text within the images, this paper proposes a high-resolution Thangka image text detection model named SAC (Sliding Window Segmentation and Classifier)-DBNet. Firstly, based on the GrabCut algorithm and the text synthesis module, high-resolution text-free images and text data crawled by Tibetan websites were automatically synthesized to expand the high-resolution Thangka image dataset. To reduce the complexity of image processing and avoid missing detection of small text size, a sliding window strategy segmentation module is added based on DBNet, and the segmented sub images are automatically labeled by coordinate mapping algorithm. Secondly, to solve the problem of resource consumption caused by the segmented image, Squeeze Net, a pre-trained lightweight model, is added as a classifier to classify whether the subgraph contains text, and the subgraph containing text is detected. Finally, according to the predicted boxes after sub image detection, they are plotted on the original image, and non-maximum suppression is used to eliminate redundant detection results. Experimental results indicate that the modified Squeeze Net classifier achieves an accuracy improvement of 3.85% over its predecessor; additionally, SAC-DBNet attains an Fl-score of 74.9%, surpassing DBNet by 8.6%, while also achieving a recall rate of 69.5%, which is an increase of 2.5%, thereby reducing missed detections in text identification and enhancing overall accuracy.