Pocket-Sized AI: Evaluating Lightweight CNNs for Real-Time Sketch Detection

Abstract

Portability with Performance Balance of neural network model is of utmost importance in the current era of Mobility. Applications and uses of deep learning neural network are growing at a fast pace and it is necessary for neural network model to be portable but capable enough to learn and identify tasks on the computational power of mobile devices. In this paper, the performance and practicability of regular Convolutional Neural Network (CNN) & depthwise separable CNN (MobileNet) are discussed and compared on mobile phones by carrying out some parametric modifications and trials on resources to study accuracy tradeoffs. Subsequently, we illustrated the implementation of depthwise separable CNN (MobileNet) on a sketch-based recognition game rewarding for successful recognition of sketch which is provided as an assignment on android platform.