A hybrid approach for plant leaf detection using ResNet50- intuitionistic fuzzy RVFL (ResNet50-IFRVFLC) classifier

Abstract

The Random Vector Functional Link (RVFL) is a prominent and widely used approach effective in tackling a wide range of challenging problems in various research fields in the case of regression and classification of real-world problems. An Intuitionistic fuzzy RVFL Classifier (IFRVFLC) boosts the overall classification performance of the RVFL network and enhances its classification accuracy on noisy datasets. On the other hand, ResNet50 architecture offers great potential in the field of artificial intelligence and is used for any object recognition task. The major drawback of ResNet50 architecture is that there is no transparency in the middle layers during the classification process, which makes it challenging to examine the training process. So, to eradicate this disadvantage, we have proposed a hybrid ResNet50-IFRVFLC model. which combines the ResNet50 and IFRVFLC models for the classification of plant species through their leaf image which is one of the biggest challenges in computer vision. Firstly, the ResNet50 model obtains the deep features using textures of the leaf images, and then PCA is applied as a feature reduction technique to extract the important features. The extracted features from PCA are taken as an input to IFRVFLC architecture for leaf image classification. The efficacy of the ResNet50-IFRVFLC model is evaluated by comparing its performance to that of the support vector machine (SVM), intuitionistic fuzzy SVM (IFSVM), twin SVM (TSVM), kernel ridge regression (KRR), RVFL, twin RVFL (TRVFL), RVFL with ε-insensitive Huber loss (ε-HRVFL) and 1-norm TRVFL (TRVFL_1norm). Furthermore, extensive statistical analysis has been performed to evaluate the significance of the noted performance differences, including Friedman and post-hoc Nemenyi tests. The experimental results are examined in terms of average accuracy, F1-Score, AUC and G-Mean. The ResNet50-IFRVFLC model achieved 91.23 % mean accuracy, outperforming SVM (88.04 %), TSVM (89.55 %), RVFL (88.84 %), TRVFL (90.31 %), IFSVM (89.47 %), KRR (87.10 %), ε-HRVFL (87.65 %) and TRVFL_1norm (90.15 %) for leaf datasets. Out of all the existing baseline models implemented in this study, ResNet50-IFRVFLC has attained the highest classification accuracy of 94.860 % and a maximum F1-score of 0.972 on the leaf datasets.