NCS based ultra low power optimized machine learning techniques for image classification

In recent years there has been an extensive development in the field of convolutional neural network-based image classification because of the human-like inference results obtained, but these massive networks are resource intensive and have high memory and computational requirements. Intel's Neural Compute Stick brings real time inference, prototyping and deployment of these DNNs to the network edge. In this paper we will discuss the development of a model for classification of book cover images into genres, and subsequently compiling the trained model for use with the Neural Compute Stick, so as to receive the optimized results in constrained environments thus ultimately leading to a system to judge a book by its cover which can be used even within a low power environment like a mobile device or Raspberry Pi, as the stick runs on power values as low as 1.2W.