Future Directions: A Journey from Handcrafted Techniques to Representation Learning
Rik Kamal Kumar Das · 2020
The popularity of deep learning techniques has initiated a paradigm shift in the domain of content-based image classification (CBIC). The entire process, from feature extraction to image classification, can now be done in an automated manner with the help of convolutional neural networks (CNNs). However, neural networks require extensive training data to ensure high classification accuracy. In smaller datasets, where training data are not in abundance, the classification results in an end-to-end CNN architecture that is not always very impressive. In such situations, feature extraction with neural network and feature fusion may result in improved performances. This chapter demonstrates the extraction of features using pretrained CNNs. It illustrates this technique with MATLAB code. The extracted features can be applied further to any standard classifier to carry out content-based image classification. This chapter also reviews the different techniques for handcrafted descriptor definition from image content discussed in the book and concludes with the future direction of such work.