ABSTRACT
This paper aims to compare the convolutional neural networks
(CNNs): ResNet50, InceptionV3, and InceptionResNetV2 tested with
and without pre-trained weights on the ImageNet database in order
to solve the scene recognition problem. The results showed that the
pre-trained ResNet50 achieved the best performance with an average
accuracy of 99.82% in training and 85.53% in the test, while the
worst result was attributed to the ResNet50 without pre-training,
with 88.76% and 71.66% of average accuracy in training and testing,
respectively. The main contribution of this work is the direct comparison
between the CNNs widely applied in the literature, that is,
to enable a better selection of the algorithms in the various scene
recognition applications.