Effects of Shortcut-Level Amount in Lightweight ResNet of ResNet on Object Recognition with Distinct Number of Categories
Aekkasit Krueangsai, Siriporn Supratid
Abstract
This paper investigates effects of shortcut-level amount in ResNet of ResNet (RoR) based upon lightweight or small-size ResNet11 on object recognition using CIFAR-100 image dataset with different sizes of 10, 50, 100 object categories. Recognition performance comparison among the traditional ResNet11, 1–and 2–shortcut-level ResNet of ResNet (1L- and 2L-RoRs) is evaluated, relying on precision, recall, F1 and accuracy scores, averaged over 10-fold cross validation. Such cross validation is performed to ensure unbiased experimental results. Confusion matrix is also considered for more detail of results investigation. The comparison results indicate ResNet11, 1L- and 2L-RoRs provide best recognition accuracy of 96.38%, 98.70% and 98.55% respectively for 10-, 50- and 100–class CIFAR-100 datasets. It is also noticed the high competency of 2L-RoR on maintaining the recognition performance as increasing the number of data classes, relative to those 1L-RoR and ResNet11. In addition, 2L-RoR model employs only 0.780% more parameters than 1L-RoR; whilst 1L-RoR utilizes a few bits 0.392% higher parameters than ResNet11.