Selected Deep Features and Multiclass SVM for Flower Image Classification

Abstract

AbstractFlower classification and recognition is an exciting research area because extensive variety of flower classes have similar colour, shape and texture features. Most of the existing flower classification systems use a combination of visual features extracted from flower images followed by classification using supervised or unsupervised learning methods. Classification accuracy of these approaches is moderate. Hence, there is a demand for a robust and accurate system to automatically classify flower images at a larger scale. In this paper, a selected deep features and Multiclass SVM based flower image classification method which uses pre-trained CNN (Convolutional Neural Network) AlexNet as feature extractor is proposed. Initially, flower image features are extracted using fully connected layers of AlexNet and subsequently most informative features are selected using minimum Redundancy Maximum Relevance (mRMR) algorithm. Finally, Multiclass Support Vector Machine (MSVM) classifier is used for classification. In the proposed scheme, computationally intensive task of training the CNN is minimized and also the efforts required to extract low level features is reduced. Classification accuracy of 98.3% and 97.7% is observed for KL University Flower (KLUF) dataset and Flower 17 dataset respectively. It is revealed that the proposed transfer learning based method outperforms existing deep learning based classification methods in terms of accuracy.KeywordsConvolutional Neural NetworkDeep learningFlower classificationSupport Vector MachineMinimum Redundancy Maximum Relevance