A Scene Text Detection Method Based on Neighbor-Level Guided and Discrete Wavelet Transformation

Within the blooming of deep learning, scene text detection has achieved great progress in these years. However, there are still challenges in scene text detection nowadays. First, the detection algorithm does not perform well in detecting different scales of instances; Second, it is difficult to detect those instances with variant aspect ratio. To handle these difficulties, a scene text detection method based on Neighbor-level Guided Network is proposed in this paper. In this paper, scene text detection network based on Neighbor-level Guided Module (NGM) and Discrete-wavelet-transform Fusion Head (DFH) is proposed. NGM is a feature fusion network contains Axial Guided Module (AGM) and General Guided Module (GGM) to adapt different semantic information in different level, making the network more robust to multi-scale and variant aspect ratio instances. To avoid incorrect text adhesion and text separation, DFH is proposed to generate accurate result. We verify the effectiveness through extensive experiments on ICDAR2015 and MSRATD-500, and commence on ablation experiments on ICDAR2015 to demonstrate the validity. Experimental results show that NGNet outperforms the baseline by 1.3% in ICDAR2015 and MSRA-TD500 in F-Measure.