A Hybrid Lightweight Deep Neural Network Approach for Plant Disease Classification Using Self-Attention Mechanism and Transfer Learning

Adem Tekerek

Makale detayları

A Hybrid Lightweight Deep Neural Network Approach for Plant Disease Classification Using Self-Attention Mechanism and Transfer Learning

Yazarlar:

Adem Tekerek

Yayıncı:

Ankara University

Ana Başlıklar:

Artificial Intelligence (Other)

Dergi:

Journal of Agricultural Sciences Vol.31 Issue 2, pp.392-412

Belge Tipi:

Makale

Yayın Tarihi:

3/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001343001

pISSN:

1300-7580

Anahtar kelimeler:

Özet:

 Classification of plant diseases is crucial for overall food security and agricultural economies in the world. However, classification has long been challenging primarily due to the various diseases it encompasses and the different environmental factors influencing them. One of the main challenges in developing an accurate classification model is obtaining high-quality, multiclass datasets. At the same time, deep learning methods like CNN may be considered state-of-the-art in detecting complex image patterns in various applications for correct diagnoses. Still, they involve poor parameter optimizations and overfitting and have very high resource requirements. This paper introduces a combined model of classifying plant diseases in an imaging approach, which incorporates the Efficient Neural Network ENN with a Squeeze and Excitation Network SEN. The architecture follows high-density feature extraction by the networks, late fusion of features, and using a cross-channel attention mechanism to boost feature representation. This work uses transfer learning to design the hyperparameter optimization scheme and early stopping scheme to avoid overfitting. We tested our model on the Plant Village Dataset and the Leaf Rose Disease Dataset with an accuracy of 96.40 for the Plant Village Dataset and 97.15 accuracy on the Leaf Rose Disease Dataset. Our model achieved higher accuracy than the traditional DNNs VGG16, Inception V3, and RESNET-50 by approximately 21.04, 9.40, and 4.33 on the Plant Village Dataset. It improved the classification accuracy compared to VGG16, Inception V3, and RESNET-50 by 15.80, 11.20, and 6.02 on the Rose leaf disease dataset, respectively. Moreover, it has the lowest times as well as space complexity: 45 minutes and 150 MB, which are less than VGG16 50 minutes, 180 MB, Inception Net 55 minutes, 170 MB, and RESNET- 50 75 minutes, 190 MB. The global results show that our approach is superior, demonstrating enhanced performance and efficiency, which makes it well-suited for real-time applications.

 Classification of plant diseases is crucial for overall food security and agricultural economies in the world. However, classification has long been challenging primarily due to the various diseases it encompasses and the different environmental factors influencing them. One of the main challenges in developing an accurate classification model is obtaining high-quality, multiclass datasets. At the same time, deep learning methods like CNN may be considered state-of-the-art in detecting complex image patterns in various applications for correct diagnoses. Still, they involve poor parameter optimizations and overfitting and have very high resource requirements. This paper introduces a combined model of classifying plant diseases in an imaging approach, which incorporates the Efficient Neural Network ENN with a Squeeze and Excitation Network SEN. The architecture follows high-density feature extraction by the networks, late fusion of features, and using a cross-channel attention mechanism to boost feature representation. This work uses transfer learning to design the hyperparameter optimization scheme and early stopping scheme to avoid overfitting. We tested our model on the Plant Village Dataset and the Leaf Rose Disease Dataset with an accuracy of 96.40 for the Plant Village Dataset and 97.15 accuracy on the Leaf Rose Disease Dataset. Our model achieved higher accuracy than the traditional DNNs VGG16, Inception V3, and RESNET-50 by approximately 21.04, 9.40, and 4.33 on the Plant Village Dataset. It improved the classification accuracy compared to VGG16, Inception V3, and RESNET-50 by 15.80, 11.20, and 6.02 on the Rose leaf disease dataset, respectively. Moreover, it has the lowest times as well as space complexity: 45 minutes and 150 MB, which are less than VGG16 50 minutes, 180 MB, Inception Net 55 minutes, 170 MB, and RESNET- 50 75 minutes, 190 MB. The global results show that our approach is superior, demonstrating enhanced performance and efficiency, which makes it well-suited for real-time applications.

 Classification of plant diseases is crucial for overall food security and agricultural economies in the world. However, classification has long been challenging primarily due to the various diseases it encompasses and the different environmental factors influencing them. One of the main challenges in developing an accurate classification model is obtaining high-quality, multiclass datasets. At the same time, deep learning methods like CNN may be considered state-of-the-art in detecting complex image patterns in various applications for correct diagnoses. Still, they involve poor parameter optimizations and overfitting and have very high resource requirements. This paper introduces a combined model of classifying plant diseases in an imaging approach, which incorporates the Efficient Neural Network ENN with a Squeeze and Excitation Network SEN. The architecture follows high-density feature extraction by the networks, late fusion of features, and using a cross-channel attention mechanism to boost feature representation. This work uses transfer learning to design the hyperparameter optimization scheme and early stopping scheme to avoid overfitting. We tested our model on the Plant Village Dataset and the Leaf Rose Disease Dataset with an accuracy of 96.40 for the Plant Village Dataset and 97.15 accuracy on the Leaf Rose Disease Dataset. Our model achieved higher accuracy than the traditional DNNs VGG16, Inception V3, and RESNET-50 by approximately 21.04, 9.40, and 4.33 on the Plant Village Dataset. It improved the classification accuracy compared to VGG16, Inception V3, and RESNET-50 by 15.80, 11.20, and 6.02 on the Rose leaf disease dataset, respectively. Moreover, it has the lowest times as well as space complexity: 45 minutes and 150 MB, which are less than VGG16 50 minutes, 180 MB, Inception Net 55 minutes, 170 MB, and RESNET- 50 75 minutes, 190 MB. The global results show that our approach is superior, demonstrating enhanced performance and efficiency, which makes it well-suited for real-time applications.

 Classification of plant diseases is crucial for overall food security and agricultural economies in the world. However, classification has long been challenging primarily due to the various diseases it encompasses and the different environmental factors influencing them. One of the main challenges in developing an accurate classification model is obtaining high-quality, multiclass datasets. At the same time, deep learning methods like CNN may be considered state-of-the-art in detecting complex image patterns in various applications for correct diagnoses. Still, they involve poor parameter optimizations and overfitting and have very high resource requirements. This paper introduces a combined model of classifying plant diseases in an imaging approach, which incorporates the Efficient Neural Network ENN with a Squeeze and Excitation Network SEN. The architecture follows high-density feature extraction by the networks, late fusion of features, and using a cross-channel attention mechanism to boost feature representation. This work uses transfer learning to design the hyperparameter optimization scheme and early stopping scheme to avoid overfitting. We tested our model on the Plant Village Dataset and the Leaf Rose Disease Dataset with an accuracy of 96.40 for the Plant Village Dataset and 97.15 accuracy on the Leaf Rose Disease Dataset. Our model achieved higher accuracy than the traditional DNNs VGG16, Inception V3, and RESNET-50 by approximately 21.04, 9.40, and 4.33 on the Plant Village Dataset. It improved the classification accuracy compared to VGG16, Inception V3, and RESNET-50 by 15.80, 11.20, and 6.02 on the Rose leaf disease dataset, respectively. Moreover, it has the lowest times as well as space complexity: 45 minutes and 150 MB, which are less than VGG16 50 minutes, 180 MB, Inception Net 55 minutes, 170 MB, and RESNET- 50 75 minutes, 190 MB. The global results show that our approach is superior, demonstrating enhanced performance and efficiency, which makes it well-suited for real-time applications.

Makale detayları

A Hybrid Lightweight Deep Neural Network Approach for Plant Disease Classification Using Self-Attention Mechanism and Transfer Learning

Yazarlar:

Adem Tekerek

Yayıncı:

Ankara University

Ana Başlıklar:

Artificial Intelligence (Other)

Dergi:

Journal of Agricultural Sciences Vol.31 Issue 2, pp.392-412

Belge Tipi:

Makale

Yayın Tarihi:

3/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001343001

pISSN:

1300-7580

Leaf classification

convolutional neural network

Efficient Neural Network

Squeeze and Excitation Network

classification accuracy

Sayı

Makale detayları

A Hybrid Lightweight Deep Neural Network Approach for Plant Disease Classification Using Self-Attention Mechanism and Transfer Learning

Yazarlar: Adem Tekerek

Yayıncı: Ankara University

Ana Başlıklar: Artificial Intelligence (Other)

Dergi: Journal of Agricultural Sciences Vol.31 Issue 2, pp.392-412

Belge Tipi: Makale

Yayın Tarihi: 3/2025

Belge Dili: İngilizce

Document ID (Org No.): 2001343001

pISSN: 1300-7580

Leaf classification

convolutional neural network

Efficient Neural Network

Squeeze and Excitation Network

classification accuracy

Sayı

Yazarlar:

Adem Tekerek

Yayıncı:

Ankara University

Ana Başlıklar:

Artificial Intelligence (Other)

Dergi:

Journal of Agricultural Sciences Vol.31 Issue 2, pp.392-412

Belge Tipi:

Makale

Yayın Tarihi:

3/2025

Belge Dili:

İngilizce

Document ID (Org No.):

2001343001

pISSN:

1300-7580