Feature Vector Learning for Images with Pretrained Models

In this paper we propose to investigate feature vector learning of images with pretrained models with an aim to classify them. We further show that clustering based on these feature vectors is much superior to that performed on raw images. We have applied fine tuning to extract features for clustering purposes. For the classification experiments, it is seen that the three pretrained models VGG19, ResNet 50 and ResNet 101 perform in a comparable manner. For the clustering with feature vectors, Silhouette score was the best for ResNet 101 Fine-tuned model for KMeans algorithm while the Calinski-Harabasz and Davies- Bouldin score was best for Agglomerative algorithm with the ResNet 101 Fine-tuned model.