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

The Random Vector Functional Link (RVFL) network provides an efficient and quick method of training feedforward single hidden neural networks. It solves major disadvantages of classical neural networks, i.e. dealing with slow convergence and overfitting by exploiting fixed random weights and closed-form output computation. DenseNet121 architecture has huge potential in the artificial intelligence field and more specifically in object recognition tasks. The major drawback, however, is its inability to provide insights into its intermediate levels of analysis of classification, which is why this algorithm offers limited flexibility to analyze the training procedure. To resolve this problem, we offer a combined solution by integrating DenseNet121 with RVFL. Within the framework described, deep features are obtained in relation to the input leaf pictures with the pre-trained DenseNet121 model that captures important texture and contour structure. To, firstly, minimize the dimension of these features and, secondly, remove superfluous redundancies, Principal Component Analysis (PCA) is used, and only the most informative components are retained. This smaller feature set is then marked to the RVFL classification, whose responsibility is a final classification. Understanding how the proposed DenseNet121-RVFL approach performs, we have conducted some comparative experiments against several baseline classifiers, such as DenseNet121- Support Vector Machine, DesnseNet121-Twin Support Vector Machine, DenseNet121-Extreme Learning Machine, DenseNet121 and DenseNet 121-Kernel Ridge Regression. The findings of the experiment reveal that the hybrid DenseNet121-RVFL model has the best effect in comparison to the other methods, as it provided the highest accuracy of 94.45, F1-Score is 0.955, Geometeric Mean(G-Mean) of 0.898 and Area Under Curve (AUC) of 0.961 on the test dataset.