Automatic Pollen Detection Based on Feature Fusion and Self-attention Mechanism

Abstract

Accurate and timely pollen prediction has great significance for pollen allergy patients. Pollen detection is one of the basic techniques for pollen prediction. In this work, we propose an automatic pollen detector based on feature fusion and self-attention mechanism, which achieves a properly balance between efficiency and precision. A simple backbone is designed to generate a relatively small model to improve the detection speed. We fuse the low-level and high-level features of smaller pollen grains to improve the detection accuracy. As pollen grains are easily broken and deformed, and usually blend with the background, thus to address these issues, we introduce the self-attention module in the proposed model to improve the detection accuracy. The experiment is operated on the pollen dataset total of 1300 images scanned by electron microscopy, and our method obtained well experimental results.