Fine-grained weed recognition using Swin Transformer and two-stage transfer learning.

Area of Science:

• Agricultural technology
• Computer vision
• Machine learning

Background:

• Weeding is essential for agriculture to prevent crop yield loss.
• Accurate weed species recognition is a key challenge for automated weeding systems.
• Visually similar weeds and crops pose difficulties for precise identification.

Purpose of the Study:

• To develop a fine-grained weed recognition method for improved accuracy.
• To enhance the ability to distinguish between visually similar weeds and crops.
• To address data limitations in training automated weed recognition models.

Main Methods:

• Utilized Swin Transformer to learn discriminative features for subtle visual differences.
• Applied contrastive loss to magnify feature distinctions between weed and crop categories.
• Implemented a two-stage transfer learning strategy to overcome insufficient training data.

Main Results:

• Achieved high recognition accuracy (99.18%), precision (99.33%), recall (99.11%), and F1 score (99.22%) on a private maize seedling and weed dataset (MWFI).
• Outperformed state-of-the-art Convolutional Neural Network (CNN) architectures like VGG-16, ResNet-50, DenseNet-121, SE-ResNet-50, and EfficientNetV2.
• Demonstrated effectiveness on the public DeepWeeds dataset, validating the proposed method's robustness.

Conclusions:

• The proposed Swin Transformer and transfer learning method significantly improves fine-grained weed recognition.
• This approach offers a robust solution for automated weed identification in agricultural settings.
• The study provides valuable insights for developing advanced automatic weed recognition systems.