https://scholars.lib.ntu.edu.tw/handle/123456789/639768
標題: | An automated online learning framework for insect pest image classification model enhancement | 作者: | Chen, Ching Kuang TA-TE LIN |
關鍵字: | Deep Learning | Gaussian Mixture Model | Insect Pest Classification | Online Learning | 公開日期: | 1-一月-2023 | 來源出版物: | 2023 ASABE Annual International Meeting | 摘要: | To effectively manage insect pests during crop growth, it is important to have accurate data on the types and numbers of insect pests present in the field. The prime way to ascertain this information is to use the image-based deep-learning recognition model for insect detection and recognition. Making a deep learning model generalizable, typically requires a large amount of manually labelled data, which can be both time-consuming and labor-intensive. Thus, this research proposed an automated online learning framework for insect pests image classification for two main purposes: (1) To automate the collection of training dataset without the need for manual labelling in order to reduce the time and effort required for data preparation; (2) To maintain or improve the model performance over time by continuously tuning the preexisting model by incorporating new data. The AIoT imaging devices developed by our previous research were used to automatically and periodically collect images of sticky paper traps. The new training samples were selected by using the Gaussian Mixture Model (GMM) trained on the features of the insect images, extracted by the CNN model, and setting the threshold by calculating the percentile value of the log-probability of the training dataset. After the new samples were collected, the fine-tuning method was used to update the previous reference model to continuously maintain or improve the model performance without human effort. In this study, the dataset was collected by our system for five years and the performance of supervised learning and the proposed automated online learning method were compared. The results showed that the proposed method gradually improved the F1-score from 0.898 to 0.956, which was comparable to the F1-score of 0.958 achieved by supervised learning. The GMM sample cleaning process ensures that the collected samples have certain feature similarities with the correct samples. Therefore, even if noisy or incorrect samples are collected, they will not cause the model to crash during the retraining process. This result indicates that the proposed method effectively enhances model performance while reducing human labor, making it a promising and resource-efficient approach for improving model performance. |
URI: | https://www.scopus.com/inward/record.uri?eid=2-s2.0-85183579046&doi=10.13031%2faim.202300292&partnerID=40&md5=dd6679e0a2ed7cde6dda39aaa2367b37 https://scholars.lib.ntu.edu.tw/handle/123456789/639768 |
ISBN: | 9781713885887 | DOI: | 10.13031/aim.202300292 |
顯示於: | 生物機電工程學系 |
在 IR 系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。