Deep Learning-Based Text Detection in Natural Scenes

Traditional text detection mainly relies on manual features, which are only applicable to simple environments and have limited generalisation capabilities. Although deep learning enhances the generalisation and robustness of detection, complex contexts still face challenges. Current CNN text detection algorithms are difficult to handle large-scale and long-distance text due to the limitation of receiving domain and spatial information extraction. This chapter proposes the GMSTNet model, which combines GhostNet V2, MobileNet V3, and Swin Transformer to enhance efficiency through segmented nonlinear activation, effectively handle small-size text and detailed features while enhancing global and local perception, and demonstrate good performance on multiple datasets.