Thin MobileNet: An Enhanced MobileNet Architecture

Abstract

In the field of computer, mobile and embedded vision Convolutional Neural Networks (CNNs) are deep learning models which play a significant role in object detection and recognition. MobileNet is one such efficient, light-weighted model for this purpose, but there are many constraints or challenges for the hardware deployment of such architectures into resource-constrained micro-controller units due to limited memory, energy and power. Also, the overall accuracy of the model generally decreases when the size and the total number of parameters are reduced by any method such as pruning or deep compression. The paper proposes three hybrid MobileNet architectures which has improved accuracy along-with reduced size, lesser number of layers, lower average computation time and very less overfitting as compared to the baseline MobileNet v1. The reason behind developing these models is to have a variant of the existing MobileNet model which will be easily deployable in memory constrained MCUs. We name the model having the smallest size (9.9 MB) as Thin MobileNet. We achieve an increase in accuracy by replacing the standard non-linear activation function ReLU with Drop Activation and introducing Random erasing regularization technique in place of drop out. The model size is reduced by using Separable Convolutions instead of the Depthwise separable convolutions used in the baseline MobileNet. Later on, we make our model shallow by eliminating a few unnecessary layers without a drop in the accuracy. The experimental results are based on training the model on CIFAR-10 dataset.