A Flower Classification Method Combining DenseNet Architecture with SVM

Abstract

Identifying flowers becomes an urgent challenge due to the light intensity at different shooting angles, complex backgrounds and similarities of different flower species. A classification method DN-F-SVM based on combining DenseNet architecture with Support Vector Machine (SVM) is proposed for flower recognition. First a convolutional architecture DenseNet with the characteristics of dense connection mechanism and feature reuse is utilized to train data. Then, flower features obtained from the connection layer of DenseNet are extracted and efficient features are selected with feature selection method of Fast Correlation-Based Filter (FCBF), which combines with SVM to achieve flower classification. The proposed classification method DN-F-SVM has been trained on the Oxford-17 and Oxford-l02 flower data sets. It has delivered recognition rates of up to 99.12% and 98.90% which are higher than the existing methods, fully demonstrating its excellent recognition performance.