Swin-HSSAM: A green coffee bean grading method by Swin transformer

Abstract

A novel shifted window (Swin) Transformer coffee bean grading model called Swin-HSSAM has been proposed to address the challenges of accurately classifying green coffee beans and low identification accuracy. This model integrated the Swin Transformer as the backbone network; fused features from the second, third, and fourth stages using the high-level screening-feature pyramid networks module; and incorporated the selective attention module (SAM) for discriminative power enhancement to enhance the feature outputs before classification. Fusion Loss was employed as the loss function. Experimental results on a proprietary coffee bean dataset demonstrate that the Swin-HSSAM model achieved an average grading accuracy of 96.34% for the three grading as well as the nine defect subdivision levels, outperforming the AlexNet, VGG16, ResNet50, MobileNet-v2, Vision Transformer (ViT), and CrossViT models by 3.86%, 2.56%, 0.44%, 4.05%, 5.36%, and 5.40% percentage points, respectively. Evaluations on a public coffee bean dataset revealed that, compared with the aforementioned models, the Swin-HSSAM model improved the average grading accuracy by 1.01%, 0.13%, 4.75%, 0.85%, 0.73%, and 0.27% percentage points, respectively. These results indicate that the Swin-HSSAM model not only achieved high grading accuracy but also exhibited broad applicability, providing a novel solution for the automated grading and identification of green coffee beans.

Fig 1. Different types of defective coffee beans.

Fig 2. Different grades of green coffee beans.

Fig 3. Image data preprocessing of Arabica green coffee beans.

Fig 4. Feature maps structure.

a The feature maps of previous ViT; b the hierarchical feature maps approach; c the shifted window approach for computing self-attention.

Fig 5. Shifted window Transformer architecture.

Fig 6. Swin Transformer block.

LN, layer norm layer; W-MSA, window-based multi-head self-attention module; MLP, multilayer perceptron. Z l, outputs of the lth W-MSA module; ẑ^l+1, shifted window-based MSA module; Z l output of the lth MLP module.

Fig 7. Architecture of the hierarchical scale-based feature pyramid network.

Fig 8. SAM discriminability enhancement module structure diagram.

Fig 9. Detailed structure diagram of the control depthwise separable convolution module.

Fig 10. Swin-HSSAM architecture.

Fig 11. Loss change curve of the Swin-HSSAM and Swin-T models(Train).

Fig 12. Loss change curve of the Swin-HSSAM and Swin-T models(Validation).

Fig 13. Loss change curve of different hierarchical scale-based feature pyramid network structure.

Fig 14. Loss change curve of different model.

Fig 15. Accuracy change curve of different mode.