ASM-DBNet: Introducing Adaptive Differentiable Binarization, Spatial-Channel Self-Attention and Multi-Scale Context-Enhanced Dynamic Upsampling for Natural Scene Text Detection

Text detection models based on DBNet have demonstrated strong performance in natural scene text detection. However, these models still suffer from the following three issues. Firstly, the amplifying factor hyperparameter in the differentiable binarization (DB) makes it difficult for the text detection model to achieve optimal performance. Secondly, the integration of low-level and high-level features within the backbone’s feature pyramid lacks specific optimization strategies. Thirdly, the deconvolution operation in the prediction head may damage text contours. To tackle the aforementioned issues, this paper presents a text detection model termed ASM-DBNet, which mainly consists of three innovations. For the first issue, an adaptive differentiable binarization (ADB) scheme is proposed. It can independently predict amplifying factor for feature points at different spatial locations and replace the original amplifying factor hyperparameter, thereby improving the overall optimization performance of the model. For the second issue, a spatial-channel self-attention (SCA) module is proposed to optimize the fusion of high-level and low-level features. On the one hand, spatial self-attention is used to enhance the spatial localization ability of high-level features; on the other hand, channel self-attention based on a grouped transformer is used to optimize the fusion results of high-level and low-level features. For the third issue, a multi-scale context-enhanced dynamic upsampling (MC-DyUpS) module is proposed to replace the deconvolution operation in the prediction head. It enhances contextual perception in the region of interpolation points through multi-scale context feature extraction, and then accurately predicts coordinate offsets of interpolation points. The position correction based on these offsets effectively suppresses the spatial deviation caused by deconvolution. Ablation studies demonstrate the effectiveness of the SCA module, MC-DyUpS module, ADB scheme, and their arbitrary combinations. Comprehensive quantitative evaluations demonstrate that ASM-DBNet achieves competitive F1-scores of 84.1%, 84.2%, and 85.7% on the ICDAR 2015, Total-Text, and MSRA-TD500 datasets, respectively, with improvements of 1.8%, 1.4%, and 2.9% over the baseline model.