Improved VGG-16 Neural Network for Parameter Reduction

Abstract

With the development of Deep Learning, image recognition technology has been applied in many aspects. And convolutional neural networks have played a key role in realizing image recognition under the increasing computing power and massive data. However, if developers want to implement the training of convolutional neural networks and achieve the subsequent applications in scenarios such as personal computers, IoT devices, and embedded platforms with low Graphics Processing Units(GPUs) memory, a large number of parameters during training of convolutional neural networks is a great challenge. Therefore, this paper uses depthwise separable convolution to optimize the classic convolutional neural network model VGG-16 to solve this problem. And the VGG-16-JS model is proposed using the Inception structure dimensionality reduction and depthwise separable convolution on the VGG-16 convolutional neural network model. Finally, this paper compares the classification success rates of VGG-16 and VGG-16-JS for the application scenario of the COVID-19 mask-wearing. A series of reliable experimental data show that the improved VGG-16-JS model significantly reduces the number of parameters required for model training without a significant drop in the success rate. It solves the GPU memory requirements for training neural networks to a certain extent.