A Comparison Study of Using Linear and Nonlinear Classifiers on Object Recognition Based on Inception Convolutional Neural Networks with Different Numbers of Inception-Block

Abstract

This paper focuses on a comparison study of using linear and nonlinear classifiers on object recognition based upon inception convolutional neural networks (ICNNs) with different numbers of inception-block. Here, the linear and nonlinear classifiers respectively refer to linear support vector classifier (LSVC) and softmax ones. The ICNNs with different sizes of 3, 6, and 9 inception-block layers, namely 3-, 6- and 9 –ICNNs extract significant features; then, individually feed them to SM and LSVC for further classification. The ICNN-SM and ICNN-LSVC, with 2-auxiliary networks are experimented. 256-resized based on unequal-sized images of Oxford-17 and Oxford-102 flower datasets, respectively having 17 and 102 classes of flower species are tested here. Recognition evaluations rely on F1 scores and accuracy rates, averaged over 10-fold cross validation to ensure unbiased experimented results. The results indicate the best F1 and accuracy scores are mostly generated by 6-ICNN based on SM and LSVC classifiers for overall results. For Oxford-17 having smaller size of dataset and classes, SM yields 18.01%, 17.21% and 16.44%better performance than LSVC for 3-, 6- and 9 –ICNN respectively; whereas for Oxford-102, a bit few differences between those classifiers are exhibited.