A hybrid DenseNet121-random vector functional link (RVFL) approach for plant leaf classification.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Classical neural networks suffer from slow convergence and overfitting.
• DenseNet121 excels in object recognition but lacks interpretability in classification.
• A need exists for improved, interpretable deep learning models in image analysis.

Purpose of the Study:

• To develop a hybrid model integrating DenseNet121 and Random Vector Functional Link (RVFL) networks.
• To enhance the classification accuracy and analytical flexibility of deep learning models for image recognition.
• To address the interpretability limitations of DenseNet121 in classification tasks.

Main Methods:

• Utilized a pre-trained DenseNet121 model to extract deep features from leaf images.
• Applied Principal Component Analysis (PCA) to reduce feature dimensionality and remove redundancy.
• Integrated the reduced feature set with an RVFL classifier for final classification.
• Conducted comparative experiments against baseline classifiers like SVM, ELM, and KRR.

Main Results:

• The hybrid DenseNet121-RVFL model achieved the highest accuracy (94.45%).
• Achieved a high F1-Score of 0.955, Geometric Mean (G-Mean) of 0.898, and Area Under Curve (AUC) of 0.961.
• Demonstrated superior performance over baseline classifiers in leaf image classification.

Conclusions:

• The proposed DenseNet121-RVFL model offers an effective and efficient solution for image classification tasks.
• Combining DenseNet121's feature extraction with RVFL's efficient classification improves performance and offers better analytical insights.
• This hybrid approach shows significant potential for advancing artificial intelligence in object recognition and similar applications.