Efficient Scene Text Segmentation via Gate Mechanism and Multi-Scale Feature Fusion

In addressing the challenges posed by current text segmentation methods when dealing with multi-scale and irregular scene text, this paper presents an innovative approach to scene text segmentation. The proposed method leverages multiple Gated Fully Fusion Modules, facilitating adaptive weight adjustments for features at various levels and selective information aggregation. This ensures enhanced capture of multiscale text information. Additionally, a Multi-Scale Feature Fusion Module further improves information capture across different scales. The incorporation of a Text Prior Module and an Enhancement Module focuses on learning text prior information and edge features, respectively, refining the precise recognition and segmentation capabilities of multi-scale text targets. Experimental evaluations on the TextSegt dataset and the Bilingual Text Segmentation dataset validate the effectiveness of the proposed algorithm. Results demonstrate outstanding performance in handling text of various languages and sizes, highlighting robust adaptability and generalization capabilities. This work presents a compelling solution for scene text segmentation tasks.