| 摘要: | 本研究開發一種創新的 CMC（CNN-Momentum Contrast-Centroid Loss）模型，結合卷積神經網路（CNN）、自監督對比學習和質心式優化（Centroid-Based Enhancement）的技術，用於玉米葉病的識別。首先，對玉米葉部的影像資料進行 RGB 強化特徵處理，以提供更豐富的視覺信息，提高模型的識別和分類能力。接著，採用 ResNet18 進行預訓練，通過對比學習學習到通用的視覺特徵表示。基於 MoCo 的動量對比學習機制，維護和更新鍵編碼器，保持查詢編碼器的特徵一致性，並結合 InfoNCE 損失函數和 Centroid Loss，形成權衡的總損失函數，進一步提升特徵空間中的類內緊密度和類間區隔性。
實驗結果顯示，CMC 模型在處理複雜和多變的農業影像資料方面具有優勢。相比於傳統的卷積神經網路和其他自監督學習模型，CMC 模型能夠更有效地識別不同類型的玉米葉病。質心式優化（Centroid-Based Enhancement）在本研究中發揮重要作用。通過引入質心式優化，CMC 模型能夠更好地學習每個類別的質心，使同類樣本在特徵空間中彼此靠近，而不同類別的樣本保持距離，從而增強類內聚合度並擴大類間差異。
為了提升 CMC 模型的性能，未來的研究將擴展資料集和病害範圍，包括增加不同生長階段和環境條件下拍攝的影像資料，並探索其他作物病害的識別。此外，跨學科合作在智慧農業中的應用，如結合空拍機、農業機器人和移動應用（APP），將提高農業監控和病害診斷的效率和精確度，推動智慧農業的發展。這些改進將提升模型的實用性和準確性，為農業監控和病害診斷提供先進的技術支持。
;This study developed an innovative CMC (CNN-Momentum Contrast-Centroid Loss) model, combining the techniques of Convolutional Neural Networks (CNN), self-supervised contrastive learning, and Centroid Loss, aimed at the identification of corn leaf diseases. Initially, RGB-enhanced feature processing was applied to the image data of corn leaves to provide richer visual information and improve the model′s recognition and classification capabilities. Subsequently, ResNet18 was used for pre-training, learning universal visual feature representations through contrastive learning. Based on MoCo′s momentum contrast learning mechanism, the key encoder is maintained and updated to keep the feature consistency of the query encoder. The combination of InfoNCE loss function and Centroid Loss forms a balanced total loss function, further enhancing intra-class compactness and inter-class separability in the feature space.
Experimental results show that the CMC model has significant advantages in handling complex and variable agricultural image data. Compared to traditional convolutional neural networks and other self-supervised learning models, the CMC model more effectively identifies different types of corn leaf diseases. Centroid-Based Enhancement played an important role in this study. By introducing Centroid Loss, the CMC model can better learn the centroid of each category, bringing similar samples closer in the feature space while keeping different categories separated, thereby enhancing intra-class aggregation and inter-class differentiation.
To further enhance the performance of the CMC model, future research will expand the dataset and the range of diseases, including images taken at different growth stages and under various environmental conditions, and explore the identification of other crop diseases. Additionally, interdisciplinary collaboration in smart agriculture, such as integrating drones, agricultural robots, and mobile applications (APPs), will significantly improve the efficiency and accuracy of agricultural monitoring and disease diagnosis, promoting the development of smart agriculture. These improvements will greatly enhance the model′s practicality and accuracy, providing advanced technical support for agricultural monitoring and disease diagnosis. |
