SPIN: Structure-Preserving Inner Offset Network for Scene Text Recognition

Arbitrary text appearance poses a great challenge in scene recognition tasks. Existing works mostly handle with the problem consideration of shape distortion, including perspective distortions, line curvature or other style variations. Rectification (i.e., spatial transformers) as preprocessing stage is one popular approach and extensively studied. However, chromatic difficulties complex scenes have not been paid much attention on. In this work, we introduce new learnable geometric-unrelated rectification, Structure-Preserving Inner Offset Network (SPIN), which allows color manipulation source data within network. This differentiable module can be inserted before any architecture to ease downstream tasks, giving neural networks ability actively transform input intensity rather than only rectification. It also serve complementary known transformations work both independent collaborative ways them. Extensive experiments show proposed transformation outperforms existing rectification has comparable performance among state-of-the-arts.