Gaussian-guided character erasure for data augment of industrial characters

The application of scene text erasure technology in privacy protection, camera-based virtual reality translation and image editing has attracted more and more research interests. Recent efforts on scene text erasing have shown promising results. We utilize text removal methods as a component of industrial characters generation procedure to generate large-scale synthetic character images so as to mitigate the issue of insufficient samples in the recognition task of industrial characters. Existing character erasure models has achieved good performance in natural scenes. However, in industrial scenes, these erasure networks are easily affected by salient no-character regions leading to the attention shift. To overcome this limitation, we proposed a character erasure network based on attention mechanism which embed an additional region awareness layer to guide attention to the correct character regions. Meanwhile, we devise a gaussian heat map supervision method for learning additional region awareness layer. The experiments show that the proposed method performs favourably on four industrial character datasets.