FSANet: A Lightweight Network for Tobacco Grouping Using Multi-scale Convolution and Attention Mechanism

To address the issues of slow classification speed, group confusion, and difficult feature extraction in the flue-tobacco grouping task using single-size convolutional kernels in traditional convolutional neural networks, we propose a lightweight network model called FSANet. This model combines a multi-scale convolutional structure and attention mechanism. Firstly, we design a multi-scale feature fusion module (FSAConv) to extract and fuse multi-scale features from flue-treated tobacco images. The module utilizes feature partitioning strategies to assign feature subsets for parallel convolution and pooling operations simultaneously, enhancing the network's sensitivity to different semantic targets. Additionally, the subset features are adaptively calibrated through the SE (squeeze and excitation) channel attention mechanism. Furthermore, FSAConv introduces the Ghost convolution module and HardSwish activation function to maintain model expressiveness while reducing parameter count. Finally, based on the inverted residual of MobileNet, we construct a lightweight network called FSANet by replacing 3 × 3 convolutions with FSAConv modules. Using our self-constructed flue-cured tobacco dataset, experimental results demonstrate that compared with GHostNet and MobileNetV3 models, FSANet exhibits an increase of 13.8% and 20% in parameter count respectively while achieving accuracy improvements of 5% and 10%. The constructed FSANet model enhances classification accuracy for flue-cured tobacco with superior overall performance suitable for actual production scenarios with limited storage resources and low hardware capabilities.