## Overview This dataset provides feature files extracted from *Tiny Imagenet Dataset*. These features are extracted through a [Python package](https://github.com/jgoodman8/py-image-features-extractor). These are the techniques used: - Convolutional Neural Network Ensemble - Linear Binary Patterns Histograms (LBPH) - Bag of Features (bag-of-visual-words) - SIFT - SURF - KAZE ## Ensemble Notice that low-resolution images make harder to get feature vectors with good classification power. By applying *feature detection* methods, it is easy to check the number of detected *key-point*s. The volume of key-points is much elevated in higher resolution samples due to their better resolution. In fact, certain images do not present any key-point for the detectors applied. Consequently, the extracted features are likely to provide poor results. In order to improve the quality of the extracted features, a new neural network has been built. The ConvNet architecture is inspired by the stacking approach by combining extracted features with different fine-tuned networks. Thus, pre-trained VGG19, ResNet50 and DenseNet201 are picked, fine-tuned over tiny epochs and their last layer is removed. As shown in the figure, data is fed into these networks and their outputs are concatenated. Finally, two dense blocks and a softmax layer are added at the top of the network. Each one of these dense blocks is built with a RELU layer, a batch normalization layer and a dropout layer. The whole model is fine-tuned and the top blocks are removed to extract high-level features.