Artificial intelligence for wildlife monitoring conservation – automatic species identification for medium to large mammals in Taiwan
- Publication Date: 2020-04-03
| Execution Methods | This study integrated image processing and deep learning models to develop automatic species identification models for medium to large mammals in Taiwan. We first developed a method to detect objects in images. We used background subtraction to extract objects (animals). We then used morphology to refine the object. We trained deep learning models to identify 14 species of animals, including humans. We trained the models with 50 images per species and validated with 15 images per species. We used confusion matrix to evaluate the performance of models. |
| Performance Evaluation | 1. Object detection method that integrates image processing can detect animals that deep learning fails to detect, especially when background is complicated or animals are covered by vegetation. 2. The deep learning models we develop can identify 14 species of mammals, including humans. The accuracy, precision, and recall rate are over 90%. 3.We applied DeepLabCut, a 3D markerless pose estimation for lab animals, to estimate position of wild macaque. We found that DeepLabCut can detect most joints of macaque. However, DeepLabCut fails to detect joints or mark wrong position when animals change direction of moving. |
| Conclusion & Suggestion | 1. Conclusion：Object detection method that integrates image processing can solve situations that deep learning fails to detect. The species identification model we trained provides foundation for future research. 2. Publication: [1] Wei-Ching Yen and Jiunn-Lin Wu*, “A Study on The Amount Prediction of The Small Size Pests in Greenhouse Condition Using Bi-Directional LSTM and Heuristic Algorithm”, The Proceedings of International Computer Symposium 2020，Tainan, Taiwan, December 17-19, 2020. [2] Kai-Ting Su and Jiunn-Lin Wu*, “A Study on The Classification of Shot Categories for The Table Tennis Game Video with The Combination of CNN and Bi-LSTM Networks”, The Proceedings of International Computer Symposium 2020，Tainan, Taiwan, December 17-19, 2020. 3. Future perspectives/ plans: Dr. Wu have applied 2021 MOST Individual Grants on deep learning and image processing. We plan to apply research grant from the Council of Agriculture in the future. |
| Appendix |