Search

To overcome the challenges posed in effectively extracting stone inscriptions characterized by highly self-similarity between the foreground and background, a character image segmentation framework is proposed that integrates Stacked-UNets and Generative Adversarial Networks (GANs). Initially, a convolutional rule tailored for self-similar feature extraction is introduced to enhance the image detail segmentation. Subsequently, in improving the stacked units, multi-scale character masks with Spatial Transformer Network (STN) are added to guide the character segmentation. Finally, with two GANs and datasets, the character recognition and generation capabilities of the Stacked-UNets are trained, and further constructing its character segmentation and restoration abilities. Ultimately, the edge segmentation performance was enhanced, both in extracting characters from stone inscriptions with self-similar blocks and in effectively recovering partially incomplete characters. Compared to state-of-the-art methods, the Stacked-UNets demonstrates improvements of 6.41%, 7.39%, 8.43%, 8.11%, 4.13%, and 2.58% across six indicators, respectively.Article highlights: The framework first establishes character image recognition and generation abilities, then, develops character image segmentation and restoration abilities. A tailored convolution operation rule is proposed to enhance the boundary segmentation performance of UNet for similar blocks. Multiple Spatial Transformer Networks are added in Stacked-UNets to extract character mask features, thereby guiding and constraining character image segmentation. [ABSTRACT FROM AUTHOR]

Abstract To overcome the challenges posed in effectively extracting stone inscriptions characterized by highly self-similarity between the foreground and background, a character image segmentation framework is proposed that integrates Stacked-UNets and Generative Adversarial Networks (GANs). Initially, a convolutional rule tailored for self-similar feature extraction is introduced to enhance the image detail segmentation. Subsequently, in improving the stacked units, multi-scale character masks with Spatial Transformer Network (STN) are added to guide the character segmentation. Finally, with two GANs and datasets, the character recognition and generation capabilities of the Stacked-UNets are trained, and further constructing its character segmentation and restoration abilities. Ultimately, the edge segmentation performance was enhanced, both in extracting characters from stone inscriptions with self-similar blocks and in effectively recovering partially incomplete characters. Compared to state-of-the-art methods, the Stacked-UNets demonstrates improvements of 6.41%, 7.39%, 8.43%, 8.11%, 4.13%, and 2.58% across six indicators, respectively.