TiTDet: A tiny text detector with scale-sensitive loss and effective fusion factor

Abstract

The text in natural scenes is often smaller compared to artificially designed text. Due to the small proportion of pixels, low resolution, less semantic information, and susceptibility to complex scenes, tiny text detection often results in many missed detections. To address this issue, this paper draws inspiration from small object detection methods and proposes TiTDet, a detection algorithm more suitable for tiny text. Due to the small proportion of pixels, low resolution, less semantic information, and susceptibility to complex scenes, tiny text detection often results in many missed detections. To address this issue, this paper draws inspiration from small object detection methods and proposes TiTDet, a detection algorithm more suitable for tiny text. Firstly, this paper incorporates a context extraction module and an attention-guided module. These modules guide contextual information learning through a self attention mechanism, while eliminating the possible negative impact caused by redundant information. Regarding multi-scale feature fusion, this paper proposes a fine-grained effective fusion factor, making the fusion process emphasize small object learning more and highlight the feature expression of tiny texts. In terms of post-processing, this paper proposes a differentiable binarization module, incorporating the binarization process into model training. Leveraging the implicit information in the data to drive model improvement can enhance the post-processing effect. Lastly, this paper proposes a scale-sensitive loss, which can handle tiny texts more fairly, fully considering the positional relationship between the predicted and real regions, and better guiding the model training. This paper proves that TiTDet exhibits high sensitivity and accuracy in detecting tiny texts, achieving an 86.0% F1-score on ICDAR2015. The paper also compares the superiority of the method on CTW1500 and Total-Text.