RGNet: A redried tobacco lamina grader network via spectral-spatial feature integration

Redried tobacco lamina (RTL) grading still relies on inefficient and subjective manual assessment. This study tackles the data scarcity and subjective grading challenges by first constructing a standardized high-quality image dataset of RTL. Building upon this foundation, an enhanced RTL recognition framework named RGNet is developed for automated grading. The architecture incorporates a Spectral-aware Dynamic Convolution (SADConv) module that performs frequency-domain decomposition and dynamic kernel generation to enhance discrimination of subtle textural patterns. Additionally, a Spatial-Channel Synergistic Bottleneck (SCSBottle) module with dual-attention structure is integrated to simultaneously model long-range spatial dependencies and channel-wise feature interactions. Trained on a dedicated high-quality dataset, our method achieves state-of-the-art performance with 95.6% classification accuracy while maintaining efficient inference speed. This demonstrates its significant potential for real-time industrial deployment. • A standardized high-quality image dataset of redried tobacco lamina addresses data scarcity and subjective grading. • A spectral-aware dynamic convolution module enhances subtle textural patterns discrimination via frequency-domain processing. • A spatial-channel synergistic bottleneck jointly models long-range spatial dependencies and channel-wise interactions. • The proposed system achieves state-of-the-art 95.6% grading accuracy and efficient real-time industrial inference.