Feature Extraction and Classification Using Deep Learning

The Image classification task is an important one in computer vision. In this work, various Deep learning methods for image classification have been extensively studied. For performing the classification and retrieval of images. Two datasets of images i.e, Celestial and Lunar data are collected to facilitate Celestial Classification for further studies. Features of these datasets are extracted using SIFT, SURF, and HOG techniques. A multi-stage CNN technique is proposed in this paper which works better for large datasets such as CIFAR 100. Its accuracy is checked for all four datasets and it is found that it works well only for either very small or very large datasets such as Celestial, Lunar, and CIFAR100. Its accuracy is compared to already existing four techniques namely Fractional Max Pooling, using Exponential Linear unit, spatially sparse CNN, and Scalable Bayesian Optimization with deep learning.