Ultra-thin MobileNet
Debjyoti Sinha, Mohamed El‐Sharkawy · 2020 10th Annual Computing and Communication Workshop and Conference (CCWC) · 2020
Convolutional Neural Networks (CNNs) are deep learning architectures which play an important role in object detection, image classification, face recognition, autonomous driving applications, etc. MobileNet is a light CNN model which was developed especially for embedded vision applications. But still, it is quite challenging to deploy the baseline model into memory constrained micro-controller units. Design Space Exploration of the above-mentioned model can make it less memory and computationally intensive. This paper proposes some modifications to the existing baseline MobileNet architecture to make it more efficient and suitable to be deployed on real-time embedded platforms. The intent behind developing such an architecture is to reduce the size, the number of parameters, computation time per epoch and the overfitting problem considerably without letting the accuracy drop below the baseline accuracy level. We achieve good accuracy levels by using the Swish activation function instead of the standard activation function ReLU and introducing a regularization method called random erasing instead of Drop out into the network. We decrease the model size by using Separable Convolutions in place of Depthwise Separable Convolutions, changing the channel depth, choosing an optimum width multiplier value and eliminating some layers with the same output shape, without much drop in the accuracy levels. We train the model with the above-mentioned modifications from scratch on the CIFAR-10 dataset and obtain a much lighter architecture as compared to the baseline MobileNet V1. We name the new DNN architecture as Ultra-thin MobileNet having a size of 3.9 MB only which is deployable in real-time embedded processors with limited memory and power.