A Relative Comparison of Different CNN Models Trained on A Dataset in The Perspective of Bangladesh

Abstract

This work evaluates modern Convolutional Neural Networks(CNN) and produces a comparative analysis to obtain the best model suitable for the distinctive scenes and objects of Bangladesh. The networks that were tested are AlexNet, ResNet50V2, ResNet152V2, InceptionV3, Inception-ResNetV2, MobileNetV2, Xception, DenseNet201. Though these models showed high performance on the huge Imagenet dataset, for colorful and highly contrasted traditional scenarios like in Bangladesh, their performances had to be compared to find the best one suited. Results have shown that DenseNet201 shows accuracy of 92.59% which is better than any other models used in this work. Xception and ResNet152V2 also performed well with an accuracy of 88.15% and 80.42%. But the accuracy drops dramatically when older models like AlexNet are implemented. For training purposes, we introduced a new dataset with the distinctive tradition in mind.