Image Classification via Semi-supervised Feature Extraction with Out-of-Sample Extension

Hierarchical Manifold Sensing with Foveation and Adaptive Partitioning of the Dataset

“…Although the training dataset consists only of 720 images, HMS performs better than the semi-supervised feature extraction algorithm in Ref. 16. Thus, for a hierarchical partitioning of the training data with k = 2 and N 1 = 1, HMS reaches an average recognition rate (over ten random splits of the data) of 94.98% with 15 sensing values.…”

“…In 2014 Dornaika et al 16 developed a semi-supervised feature extraction with an out-of-sample extension algorithm which they applied on a subset of the COIL-20 (18 images from 72 available for each object) database. They randomly selected 50% of the data as the training dataset and the rest as the test dataset.…”

“…The work Ref. 16 provides a comparison between methods that are based on label propagation and on graph-based semi-supervised embedding. They report the best average classification results on ten random splits for their method for three label samples and for unlabeled (80.4%) and test data (77.4%).…”

“…In order to compare HMS with the approach proposed in Ref. 16 we divided the COIL dataset with 72 objects per class into training and test datasets in a similar way to that described before. Although the training dataset consists only of 720 images, HMS performs better than the semi-supervised feature extraction algorithm in Ref.…”