FSANet: Feature shuffle and adaptive channel attention network for arbitrary shape scene text detection

Natural scene text detection has made significant progress in the era of deep learning. However, existing methods still exhibit deficiencies when faced with challenges such as complex backgrounds, extreme aspect ratios, and arbitrary-shaped text. To address these issues, we propose a segmentation-based Feature Shuffle Attention Network (FSANet) designed to enhance high-resolution feature extraction and multi-scale feature enhancement for robust detection of arbitrary-shaped text. FSANet is composed of two principal modules: (1) the High-Resolution Feature Extraction Network (FEN), which employs two Group Shuffle Blocks (GSBs) to maintain high-resolution details and promote feature interaction and information flow, and (2) the Adaptive Channel Attention Module (ACAM), which reduces background noise and redundant features by adaptively learning inter-feature correlations across scales, assigning weights to prioritize local features within a global context. Extensive experiments conducted on four public benchmark datasets show that, compared to the baseline method, the proposed method demonstrates an improved F-measure on the ICDAR2015, Total-Text, MSRA-TD500, and ICDAR2017-MLT datasets, with an average increase of 1.68%. The recall metric also consistently improves, with an average increase of 2.0%. Notably, the proposed method achieves the highest F-measure of 85.4%, 75.9% on the ICDAR2015, ICDAR2017-MLT datasets, respectively. Furthermore, on the other two datasets, the performance of the proposed method surpasses that of most existing methods, indicating that FSANet outperforms the majority of state-of-the-art approaches. The code will be publicly released at https://github.com/runminwang/FSANet.