基于频谱增强与跨域注意力的药品包装文本检测方法OA

To address complex interference challenges in pharmaceutical packaging text detection within smart healthcare scenarios,such as surface reflections,dense tiny characters,and non-planar deformations,a detection method based on spectral enhancement and cross-domain attention,named ArcCRAFT,is proposed.A dynamic receptive field adaptation mechanism was designed,employing a gradient-guided dilation rate routing strategy to enhance the model's perception of multi-scale broken text.A spectral feature enhancer was constructed,utilizing discrete Fourier transform to map spatial features to the frequency domain,suppressing high-frequency reflective noise while strengthening text edge features.A cross-dimensional global attention mechanism was introduced,enabling parallel interaction of channel,spatial,and pixel-level features to improve the analysis capability for complex textures.Experimental results on the PharmaBox dataset demonstrate that the ArcCRAFT model achieves a precision of 91.1%,a recall of 91.0%,and an F1-score of 91.0%,outperforming advanced methods such as ISTD-DLA and PP-OCRv5-server.This research provides a highly robust technical solution for automated dispensing in smart pharmacies.