Impacts of Kernel Factorization on Different Resized Images in Object Recognition Base on Inception Convolutional Neural Network

Abstract

This paper investigates impacts of kernel factorization on different resized images in object recognition based on three inception-block modules of inception convolutional neural networks (ICNNs), namely ICONV3 and ICONV131. The ICONV3 employs 3×3 kernel-size convolutions according to each inception-block; whereas the ICONV131 equivalently uses 1×3 and 3×1, replacing one 3×3 convolutions. The experiments rely on two different resized 50×50, 100×100 pixels regarding 60,000 samples of 32×32, in 10-class benchmark CIFAR-10 image dataset. Recognition performance evaluations depend on averages of precision, recall, F1 and accuracy scores, based on 10-fold cross validation for bias reduction purpose. The results indicates approximate 5% obviously better recognition accuracy of using larger resized of CIFAR-10 than small ones according to both ICONV3 and ICONV131 models. Only few 0.2% to 0.3% different recognition performance are exhibited between ICONV3 and ICONV131 for both large and small resized images. This may cause one to say ICONV131 may be preferred to ICONV3 due to about 3.6% and 6% greater parameters used in ICONV3 than ICONV131, respectively regarding 50×50 and 100×100 resized images.