VGG16 feature selection using PCA-big bang big algorithm

Abstract

In the recent decade, plant disease classification using convolution neural networks has proven to be superior because of its ability to extract key features. Obtaining the optimum feature subset with the necessary discriminant information is challenging. The main objective of this paper is to design an efficient hybrid plant disease feature selection approach and validate it on standard image datasets. The raw input image features were transformed into 8192 learned features by employing the VGG16. To reduce the training time and enhance classification accuracy, the dimensionality reduction technique Principal Component Analysis (PCA) is integrated with the big bang-big crunch (BBBC) optimization algorithm. The PCA-BBBC feature selection method reduces computing time by eliminating unnecessary and redundant features. The proposed approach was evaluated on plant diseases and benchmarked image datasets. Experimental results reveal that the Artificial Neural Network (ANN) classifier integrated with the VGG16-PCA-BBBC approach enhanced the performance of the classifier. The proposed approach outperformed the VGG16-PCA-ANN method and other popular image classification techniques. For the rice disease dataset, the proposed hybrid approach reduced the VGG16 extracted 8192 deep features to 200 relevant principal components. The recommended reduced features were used for training ANN. The test dataset was classified by ANN with an accuracy of 99.12%. Experimental results demonstrate that the proposed approach improved the performance of the classifier and accurately labeled image and plant diseases datasets aiding farmers to adopt remedial measures.