Performance evaluation of adaptive pruning algorithm for CNN accelerated vehicle re-identification in smart city environment

Abstract

Smart cities are rising as solution to many issues caused by rapid growths in urbanization and urban population. The video surveillance systems play an important role in smart city areas. Among many DL-based applications employed by the surveillance systems, the vehicle re-identification is highly important in security and traffic control. In the work of re-identification algorithm development, many existing methods aim to accelerate the convolutional neural networks (CNNs), that indirectly accelerate the CNN-based vehicle re-identification systems. They typically exploit the neural network pruning because of its competitive performance and compatibility. However, a common limitation of previous approaches is that they only determine the pruning degree of network layers individually, leading to a modest efficiency in the acceleration. In this thesis, we overcome this limitation by developing an adaptive pruning scheme for CNN acceleration. Our scheme focuses on calculating the pruning degree across multiple network layers, instead of individual layers as the old methods. It can operate in an automatic manner and allow to aggressively accelerate the CNNs in vehicle re-identification systems. The pruned models produced by our scheme are applied in vehicle re-identification systems of smart city environments. In terms of evaluation, we conducted two experiments. The first experiment was to compare our pruned model performance with the previous method. We utilized the VGG16 trained on CIFAR10 dataset as the baseline, and set up three different target computation saving costs of 25%, 50%, and 75%, respectively. The experimental results showed that our scheme obtained a higher accuracy while saving the comparable amount of computation cost. In the second experiment, we applied the pruned models produced by our scheme to a vehicle re- identification system in the smart city environment. We employed two CNN model architectures (i.e. VGG Net and Residual Net) and utilized different pruning degrees (from 10% to 60%) to verify the applicability of our proposed approach. The results proved our proposed scheme was able to save up to 50% of execution time in the CNN feature extraction module, while maintaining a desired accuracy of the overall system.