UFNet: A Multi-scale Fusion Feature based Text Detection Method

Recently, the field of text detection has witnessed a growing trend, with more and more segmentation-based methods incorporating feature sampling. Segmentation methods possess a natural advantage in detecting text with both regular and irregular shapes due to their ability to effectively segment diverse targets and backgrounds that exhibit significant differences.The common sampling method in networks is typically the Feature Pyramid Network (FPN), which is used to match different dimensions for detecting the scale of images. However, due to the inherent limitations of scene text, such as variations in aspect ratio, dense text, and differences in width-to-height ratio, these general sampling methods (FPN) may not effectively address these issues. To ease this problem, we have proposed a novel network architecture called Unified Feature Fusion Network (UFNet), which integrates feature sampling. Compared to the DBU network, UFNet achieves significantly better performance in terms of accuracy and recall on English text detection datasets such as ICDAR2015 and the mixed English-Chinese dataset MSRA-TD500. Text detection results indicate that this algorithm solves the problem of poor performance in handling variations in aspect ratio and width-to-height ratio in images.