Research on image segmentation based scene text detection algorithm

Aiming at the problems of over-segmentation and text adhesion in natural scene text detection algorithms based on image segmentation， a natural scene text detection algorithm based on a cross-level attention mechanism is proposed. By designing a cross-level attention module， the network's focus on key features and contextual information in high-resolution feature maps is enhanced by applying the proposed algorithm that thereby improves the integration capability for fragmented text. Through the design of a feature decomposition and reorganization module， the fused features are decomposed into high-frequency and low-frequency components that enhance the network's ability of distinguishing text boundary regions. After integrating these two modules into the baseline model， performance tests are implemented on two mainstream datasets. The experimental results show those of current mainstream algorithms are all surpassed， and compared to the baseline model， the missed detection rate and false alarm rate are both declined.