AirNeXt and Centroid-Based Loss Adjustment Method for Agricultural Image Detection : Case Studies on Chili Seed Germination and Coffee Leaf Disease = 농업 이미지 탐지를 위한 AirNeXt 및 중심점 기반 손실 함수 조정 기법: 고추 종자 발아와 커피잎 병해 사례 연구

다국어 초록 (Multilingual Abstract)

With the in-depth application of artificial intelligence technology in the field of smart agriculture, effectively balancing the recognition accuracy and computational efficiency of deep learning models in resource-constrained practical deployment scenarios has become a core challenge. To address this challenge, we propose and systematically validate a comprehensive solution from two complementary dimensions: model architecture design and training process optimization. In this dissertation, we design a new lightweight convolutional neural network architecture called AirNeXt. Through the innovative residual grouped convolution and multi-scale feature fusion module, we significantly reduce the number of model parameters and memory consumption and show comparable or even better performance than the heavyweight baseline model in multiple agricultural image classification tasks. To enhance the model's feature discrimination ability in complex classification tasks, this dissertation proposes a novel centroids-based loss adjustment method. This method effectively guides the model to learn more discriminative feature distributions by intelligently adjusting sample weights during training, without increasing the model's inference overhead. And it significantly improves the classification performance of multiple mainstream CNN architectures. Through the verification of two specific cases—pepper seed germination status classification and coffee leaf disease identification —the final synergy effect experiment demonstrates that combining the AirNeXt model with the centroid-based loss adjustment method achieves the best performance-efficiency trade-off on all test tasks. Especially in the real-world dataset with class imbalance, the collaborative combination scheme achieves the highest F1 score, which shows its excellent robustness and practical value. In summary, the results of this study show that a holistic strategy for collaborative optimization of accuracy and efficiency is an effective way to promote the landing application of advanced AI technology in the field of smart agriculture.

목차 (Table of Contents)

• List of Figures iv
• List of Tables v
• List of Abbreviations vi
• (Abstract) vii
• 1. Introduction 1
• 1.1 Problem Statement 1
• 1.2 Research Purpose 3
• 1.3 Contributions 5
• 1.4 Dissertation Outline 7
• 2. Background and Related Works 8
• 2.1 Smart Agriculture and Computer Vision 8
• 2.2 Key Challenges in the Agricultural Scenarios 9
• 2.2.1 Challenges in Non-destructive Detection of Chili Seed Germination 9
• 2.2.2 Challenges in Coffee Leaf Disease Identification and Diagnosis 12
• 2.3 Mainstream Methodologies for Agricultural Image Analysis 15
• 2.3.1 Research on Lightweight Deep Learning Models 15
• 2.3.2 Optimization Strategies for Deep Learning Loss Functions 17
• 3. Chili Seed Germination Classification 20
• 3.1 Introduction 20
• 3.2 Materials and Methods 21
• 3.2.1 Dataset Description 22
• 3.2.2 Data preprocessing 26
• 3.3 AirNeXt 28
• 3.4 Experiment 33
• 3.4.1 Experimental Setup and Hyperparameter Tuning 33
• 3.4.2 Model Performance Comparison 37
• 3.4.3 Background Removal Experiment Results 40
• 3.4.4 Ablation and Variant Experiment Results for AirNeXt 42
• 3.4.5 Confusion Matrix Analysis 44
• 3.4.6 Heatmap Visualization and IoU Evaluation 46
• 3.4.7 Discussion 49
• 3.5 Conclusion 51
• 4. Coffee Leaf Disease Classification 52
• 4.1 Introduction 52
• 4.2 1 Materials and Methods 54
• 4.2.1 Dataset Description and Preprocessing 55
• 4.2.2 Backbone 60
• 4.2.3 Loss Function 61
• 4.3 Centroid-Based Loss Adjustment 62
• 4.4 Experiments 65
• 4.4.1 Experimental Setup 65
• 4.4.2 Baseline Model Performance 66
• 4.4.3 Effectiveness of Centroid-Based Loss Adjustment 71
• 4.4.4 Comparison with State-of-the-Art Approaches 73
• 4.4.5 Computational Efficiency and Training Time 76
• 4.4.6 Discussion 77
• 4.5 Conclusion 78
• 5. Synergy Experiments 80
• 5.1 Introduction 80
• 5.2 Experimental setup 81
• 5.3 Results and Analysis 82
• 5.4 Conclusion 85
• 6. Conclusion and Future Works 87
• 6.1 Conclusion 87
• 6.2 Future Works 88
• References 91
• Korean Abstract 100
• Acknowledgment 102