Related Experiment Video
Updated: Jul 21, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
571
Detection of Hindwing Landmarks Using Transfer Learning and High-Resolution Networks
Yi Yang1,2, Xiaokun Liu1,2, Wenjie Li1,2
1Key Laboratory of Zoological Systematics and Evolution, Institute of Zoology, Chinese Academy of Sciences, Beijing 100101, China.
Biology
|July 29, 2023
Summary
This study introduces a new method for automatically detecting landmarks on leaf beetle hindwings, significantly speeding up insect wing analysis. The approach achieves high accuracy even with minimal training data, outperforming existing methods.
Area of Science:
- Entomology
- Morphology
- Computer Vision
Background:
- Hindwing venation is crucial for analyzing beetle flight and designing bio-inspired micro aerial vehicles.
- Manual landmark annotation for hindwing morphology is labor-intensive and limits research progress.
Purpose of the Study:
- To develop an automated method for detecting landmarks on leaf beetle hindwings.
- To enable efficient wing morphology analysis using limited sample data.
Main Methods:
- Utilized a deep high-resolution network pre-trained on a large natural image dataset (ImageNet).
- Transferred the model to the leaf beetle hindwing domain by retraining high-stage network parameters.
- Constructed a leaf beetle hindwing landmark dataset for training and evaluation.
Main Results:
- Achieved an average detection normalized mean error below 0.02 with 100 training samples.
- Maintained an error of only 0.045 with as few as three training samples.
- Demonstrated superior performance compared to a deep residual network approach.
Conclusions:
- The proposed model transfer approach is practical for leaf beetle hindwing landmark detection.
- Leveraging natural image datasets for pre-training offers a promising avenue for insect wing venation digitization.
- This method accelerates morphological analysis and supports advancements in entomology and bio-inspired engineering.
