Authors:
Matheus Todescato
;
Luan Garcia
;
Dennis Balreira
and
Joel Carbonera
Affiliation:
Institute of Informatics, UFRGS, Porto Alegre, Brazil
Keyword(s):
Image Classification, Transfer Learning, Deep Learning, Geology.
Abstract:
Dealing with image retrieval in corporate systems becomes challenging when the dataset is small and the images present features in multiple scales. In this paper, we propose the notion of multiscale context features, in order to decrease information loss and improve the classification of images in such scenarios. We propose a preprocessing approach that splits the image into a set of patches, computes their features using a pre-trained model, and computes the context feature representing the whole image as an aggregation of the features extracted from individual patches. Besides that, we apply this approach in different scales of the image, generating context features of different scales, and we aggregate them to generate a multiscale representation of the image, which is used as the classifier input. We evaluated our method in a geological images dataset and in a publicly available dataset. We evaluate our approach with three efficient pre-trained models as feature extractors. The e
xperiments show that our approach achieves better results than the conventional approaches for this task.
(More)