Performance comparison of AlexNet and MobileNetV2 architectures in flower classification

Abstract

Due to the high diversity, the problem of classifying flower species is a difficult and costly process. Computer vision and deep learning applications provide great advantages to facilitate the work of researchers working in this field. With the development of new algorithms, deep learning methods can achieve high success in performing fast and accurate analysis on large amounts of data. These methods, which are used in many fields, also provide successful results in classifying flower species. Oxford-17 data set was used in this study. The data set includes 1360 flower images belonging to 17 classes. In this study, which was created using convolutional neural networks, AlexNet and MobileNetV2 architectures, which are deep learning architectures, were used. 4096 features were extracted using AlexNet architecture. By extracting 4096 features from each image, a feature matrix of 1360x4096 size was obtained. The first 500 features were selected with the Neighborhood Component Analysis (NCA) algorithm and the classification process was carried out with the Support Vector Machine (SVM) method. Then, 1000 features were extracted using the MobileNetV2 architecture and a feature matrix of 1360x1000 size was obtained. By repeating the same processes, the first 500 features were selected using the NCA algorithm and classification was performed with SVM. By comparing the performance of the two architectures, a success rate of 93.1% was obtained from the AlexNet architecture and 93.9% from the MobileNetV2 architecture. It has been observed that the result obtained from MobileNetV2 architecture is more effective.