DEVELOPMENT OF A CONVOLUTIONAL NEURAL NETWORK ARCHITECTURE FOR MILITARY EQUIPMENT IMAGE CLASSIFICATION ON A LIMITED DATASET

Abstract

The work is aimed at developing the architecture of a convolutional neural network for the classification of military equipment images. The key requirement for the model is the ability to be trained on limited data samples. The existing VGG-16 model was chosen, in which fully connected layers were replaced with a classifier based on a fully connected neural network with 2 outputs, resulting in a model with 13 convolutional blocks with a maximizing aggregation layer between each and 3 fully connected layers. Each convolutional layer was pre-trained using the ReLU activation function. Two fully connected layers of the replaced classifier also use ReLU as an activator, the last one-node layer uses a sigmoid function to perform the classification. In order to prevent retraining of the network, the Dropout regularization method with a screening factor of 0.2 was applied. To train the convolutional neural network, we use the Normal vs Military Vehicles dataset. The size of this dataset is quite limited and consists of approximately 17,500 files divided into training, verification and testing sets, each being divided in two asset classes: military and non-military. To overcome this limitation, the transfer learning method was used, when elements of a previously trained model are reused in a new machine learning model. Since VGG-16 is trained on a very large image set and is specially designed for image recognition and classification, resulting models show high accuracy and performance even after being trained on a limited datasets. In our case showing a decent 82% accuracy on Normal vs Military Vehicles validation dataset. All trained models are checked for adequacy, and will be used in the future for quick identification and classification of military equipment in the video stream.