Impacts of ResNet Skip Connection levels on Inception Convolutional Neural Network using Different Resized Images in Object Recognition
Tatpon Kongkertsuk, Siriporn Supratid · 2023
This paper focuses on impacts of using 1 -and 2 - level ResNet skip connection on inception convolutional neural network (ICNN), named here as 1L - and 2L -ICRN in object recognition. The 1L- and 2L -ICRN as well as ICNN are brought into comparison studies using CIFAR-10 image dataset, with 70x70, 90x90 and 110x110 resized images. Recognition performance appraisements count on averages of F1, accuracy scores, recall and precision, relying upon 5-fold cross validation for bias reduction purpose. Confusion matrix is also examined for more detail of results inspection. The results denote that 1L-ICRN yields 83.02%, 84.85%, 85.06% best recognition accuracy based on 70x70, 90x90 and 110x110 images, consecutively. However, using 70 × 70 images, 1L-ICRN generates 1.05% and 1.68% more accuracy than 2L-ICRN and ICNN, respectively. As image size increases, 1L - and 2L -ICRN generate better performance but in decreasing rate; whilst, ICNN exhibits decreasing performance when expanding the size from 90 X 90 to 110 X 110. Nevertheless, at most, 2-second difference of time consumed by each model is pointed, which is insignificant.