Feature Fusion and Augmentation based on Manifold Ranking for Image Classification

Abstract

Despite the great advances in the field of image classification, the association of ideal approaches that can bring improved results, considering different datasets, is still an open challenge. In this work, a novel approach is presented, based on a combination of compared strategies: features extraction for early fusion; rankings based on manifold learning for late fusion; and feature augmentation applied in a long short-term memory (LSTM) algorithm. The proposed method aims to investigate the effect of feature fusion (early fusion) and rankings fusion (late fusion) in the final results of image classification. The experimental results showed that the proposed strategies improved the accuracy of results in different tested datasets (such as CIFAR10, Stanford Dogs, Linnaeus5, Flowers102, and Flowers17) using a fusion of features from three convolutional neural networks - CNNs (ResNet152, VGG16, DPN92) and its respective generated rankings. The results indicated significant improvements and showed the potential of the approach proposed for image classification.