Improving a 3-D Convolutional Neural Network Model Reinvented from VGG16 with Batch Normalization

It is challenging to build and train a Convolutional Neural Network model that can achieve a high accuracy rate for the first time. There are many variables to consider such as initial parameters, learning rate, and batch size. Unsuccessfully training a model is one of the most inevitable problems. In some cases, the model struggles to find a lower Loss Function value which results in a poor performance. Batch Normalization is considered as a remedy to overcome this problem. In this paper, two models reinvented from VGG16 are created with and without using Batch Normalization to evaluate their model performance. It is clear that the model using Batch Normalization provides a better result in terms of Loss Function value and model accuracy, which also achieves a very high accuracy rate. It also reaches the saturation point of the highest model accuracy faster than the model without Batch Normalization. This paper also finds that the accuracy of 3D Convolutional Neural Network model reinvented from VGG16 with Batch Normalization is at 91.2% which can beat many benchmarking results on UCF101 such as IDT [6], Two-Stream [11], and Dynamic Image Networks IDT [4]. The technique introduced in this paper shows a fast, reliable and accurate estimation of human activity type and could be used in smart environments.