Military Vehicle Recognition with Different Image Machine Learning Techniques

Abstract

Different neural network training systems are studied for image recognition of military vehicles, variable start layer transfer training models and own convolutional neural networks training from scratch. Since, there is limited openly available military recordings, labeled social media images are used for training. Furthermore, expanding the image-set by random data transformation. An implementation is made in terms of image augmentation handling as an internal loop that freezes all numerical parameters of the neural network training, while selecting continuously a slightly larger section of the training set including an increment part of artificial images added to the system. All models where trained for three vehicle and two situational environment classification cases. The transfer learning is based on two of the most widely used recognition networks, ResNet50 and Xception, with a variable number of last trained layers to max. twenty. The first being successfully transfer-trained with validation accuracy values of ≈ 88%. In contrast Xception resulted on a over-fitted neural network with low validation accuracy and large loss values. Neither of the transferred schemes benefit from image augmentation. Moreover, in variable architecture training of convolutional networks, it was corroborated that different configurations of layers numbers/type/neurons adapt differently. Thus, a tailor-fit neural network combined with data augmentation strategy is the best approach with validation accuracy of ≈ 86.4%, comparable to large transferred networks with a ≈ 40 times smaller network architecture. Hence, requiring less computational resources. Data augmentation influenced an increment of validation accuracy values of ≈ 9.2%, with the least accurate network trained gaining up to 20% on accuracy due inclusion of artificial images.