Investigating Deep Discriminative Features for Plant Disease Detection with Deep Learning Sana Parez = 식물 질병 탐지를 위한 딥러닝 기반 심층 판별 특징 분석에 관한 연구

다국어 초록 (Multilingual Abstract)

A country’s economic growth heavily depends on agricultural development; however, crop growth rates and quality are severely impacted by various plant diseases. The accurate identification of these diseases remains challenging and time-consuming due to the scarcity of domain experts and low-contrast information. To address this issue, the agricultural management system seeks automated methods for early disease detection. In response to these challenges, in this thesis two efficient deep learning (DL) models are proposed for plant disease detection in precision agriculture. First, we present E-GreenNet, a lightweight deep learning framework tailored to overcome the limitations of manual disease identification. E-GreenNet adopts the MobileNetV3Small model as a backbone, generating refined, discriminative, and prominent features. The proposed model is trained on three distinct datasets, including plant village (PV), data repository of leaf images (DRLI), and a new plant composite (PC) dataset. Extensive experimental analysis reveals impressive accuracies of 1.00%, 0.96%, and 0.99% on PV, DRLI, and PC datasets, respectively. Additionally, E-GreenNet demonstrates superior inference speed compared to other state-of-the-art approaches, making it a powerful tool for efficient and accurate plant disease classification. Next, we explore GreenViT, a fine-tuned technique based on vision transformers (ViT) for automated plant disease detection. Leveraging ViT’s strengths, GreenViT divides input images into smaller blocks or patches and feeds them sequentially to the ViT, overcoming limitations associated with traditional convolutional neural network (CNN)-based models. Through experiments on widely used benchmark datasets, GreenViT outperforms CNN models, showcasing its effectiveness in detecting plant diseases accurately and efficiently. By introducing E-GreenNet and GreenViT, we contribute towards sustainable agricultural systems by facilitating efficient and accurate plant disease detection. These models alleviate the burden of manual identification, enabling improved crop management and enhanced agricultural productivity. The combination of E-GreenNet lightweight architecture and GreenViT fine-tuned approach establishes a comprehensive framework for early-stage disease detection, fostering precision agriculture and driving economic growth in the agricultural sector. In conclusion, our proposed deep learning models demonstrate their potential to revolutionize plant disease detection in precision agriculture, offering powerful tools to address the challenges posed by plant diseases and pave the way towards sustainable agricultural practices. Keywords: agriculture monitoring; computer vision; deep learning; embedded vision; classification; plant disease detection; precision agriculture; internet of things (IoT); vision transformers.

목차 (Table of Contents)

• Abstract i
• List of Figures vi
• List of Tables viii
• List of Publications x
• 1 Introduction 1
• 1.1 Overview 1
• 1.2 Problem Definition . 5
• 1.3 Thesis Organization . 6
• 2 Literature Review 8
• 2.1 Overview 8
• 2.2 Machine Learning and Deep Learning Approaches 8
• 2.2.1 CNN-based Plant Disease Detection 11
• 2.2.2 Vision Transformers based Plant Disease Detection . 13
• 3 Efficient Deep Learning Methods for Plant Disease Detection 16
• 3.1 Overview 16
• 3.2 Architecture Details of E-GreenNet . 16
• 3.2.1 Depth-wise Separable Convolution . 18
• 3.2.2 Linear Bottleneck . 18
• 3.2.3 Inverted Residual . 19
• 3.2.4 Network Architecture Search . 20
• 3.2.5 Swish Function 21
• iii
• 3.3 Architecture Details of GreenViT 21
• 3.3.1 Embedding Layer . 22
• 3.3.2 Encoding Layer 24
• 3.3.3 Classification Layer 25
• 4 Experimental Results and Discussion 27
• 4.1 Overview 27
• 4.2 Datasets . 27
• 4.2.1 Plant Village 28
• 4.2.2 Data Repository of Leaf Images 28
• 4.2.3 Plant Composite . 29
• 4.3 Experimental Setup . 30
• 4.4 Evaluation Metrics . 31
• 4.5 Experimental Evaluation of E-GreenNet 31
• 4.5.1 Performance Comparison with State-of-the-art Methods . 32
• 4.5.2 Time Complexity Analysis 35
• 4.6 Experimental Evaluation of GreenViT 38
• 4.6.1 Performance Comparison with State-of-the-art Methods . 38
• 4.6.2 Performance Comparison with State-of-the-art Methods . 41
• 4.6.3 Qualitative Analysis 43
• 4.6.4 Time Complexity Analysis 44
• 5 Conclusion and Future Research Direction 47
• 5.1 Overview 47
• 5.2 Future Research Directions 48
• References 51
• iv
• 국문초록 64
• 감사의글 66
• v