Efficient agricultural pest classification using vision transformer with hybrid pooled multihead attention.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate pest classification is crucial for effective agricultural pest management and crop yield.
• Convolutional Neural Networks (CNNs) struggle with capturing both local and global image features.
• Vision transformers offer improved global dependency capture but traditional attention mechanisms limit fine-grained detail analysis.

Purpose of the Study:

• To introduce a novel Hybrid Pooled Multihead Attention (HPMA) model for superior pest classification.
• To overcome the limitations of existing CNNs and vision transformers in capturing spatial relationships and fine-grained details.
• To enhance the robustness and generalization capabilities of pest identification models.

Main Methods:

• Development of the Hybrid Pooled Multihead Attention (HPMA) model integrating hybrid pooling and modified attention mechanisms.
• Training and testing the HPMA model on a new dataset of 10 pest classes.
• Validation of the HPMA model on two established benchmark datasets.

Main Results:

• The HPMA model achieved a remarkable 98% accuracy on a newly created 10-class pest dataset.
• Validation on benchmark datasets yielded accuracies of 98% and 95%, demonstrating broad effectiveness.
• The model effectively captures local and global features, emphasizing discriminative characteristics for improved pest identification.

Conclusions:

• The proposed HPMA model significantly outperforms traditional CNNs and standard vision transformers in pest classification tasks.
• HPMA's ability to capture both local and global image information leads to heightened accuracy and robustness in pest identification.
• This advancement in pest classification technology aids in prompt pest control, reducing crop losses and enhancing agricultural productivity.