Effects of Number and Position of Auxiliary Networks used in Inception Convolutional Neural Network on Object Recognition
Satawat Singprayoon, Siriporn Supratid · 2021
This paper focuses on effects of number and position of auxiliary networks used in inception convolutional neural networks (ICNNs) on object recognition. The ICNNs, using six modules of inception- block having two-auxiliary network (2-ICNN), one- auxiliary-network-after-first-inception-block (1F- ICNN), one-auxiliary-network-before-last-inception- block, (1L-ICNN) are experimented here. According to 2- ICNN, two auxiliary networks are inserted after the first inception-block and before the last one; whilst the auxiliary networks before the last inception-block and after the first one are removed according to 1F-ICNN and 1L-ICNN, consecutively. The experiments rely on Oxford-17 and Oxford-102 flower datasets. Recognition performance assessments depend on averages of F1 and accuracy scores, based on 10-fold cross validation for bias reduction purpose. The results indicate that 1F-ICNN yields 81.88% and 86.39% best recognition performance for Oxford-17, containing 17 classes of flower species; whereas, 2-ICNN provides 70.16% and 79.29% best performance for Oxford-102 with 102 classes of species, based on 70 × 70 and 140 × 140 pixels resized images.