Analyzing CNN Performance for Image Processing Tasks

This research conducts a performance evaluation of five established CNN architectures -- ResNet, AlexNet, MobileNet, VGG16, and GoogleNet -- on the CIFAR-100 dataset, all trained under the same conditions. ResNet had the highest test accuracy of 75%, followed by GoogleNet (73.1%), VGG16 (58.99%), MobileNet (40%), and AlexNet (40%). The images per inference time varied from 2 ms (ResNet) to 31 ms (GoogleNet) with the parameters varying from 4.2M (MobileNet) to 138M (VGG16). Overall the results showed the trade-offs between accuracy and computational cost with useful benchmarks for model choices in resource constrained applications.