Efficient Vehicle Type Classification with Distinctiveness-based CNN Pruning
Final project for the course COMP4660 - Neural Networks, Deep Learning and Bio-inspired Computing at the Australian National University.
Abstract
As one of the essential parts of the modern AI city system, Vehicle type classification (VTC) aids a wide range of applications like vehicle re-identification and traffic flow estimation. Many existing methods are based on convolutional neural networks (CNNs), a deep learning method that can recognize vehicle types quickly and precisely. However, current CNNs usually have a large number of computing layers, which makes it hard to be deployed on a mobile system like the street camera. In this paper, we propose distinctiveness-based pruning, a structured pruning method that can effectively remove redundant filters from the convolutional layers. We test the pruning method on a classic CNN architecture VGG-11. The extensive experiment conducted on synthetic dataset VehicleX shows that the distinctiveness-based pruning can reduce 56% of the network parameters with only 0.25% test accuracy lost. We can also prune up to 88% of the parameters with only 7.5% relative test accuracy lost. Our results also suggest that distinctiveness-based pruning is more effective than the classic structured CNN pruning algorithms in reducing the size of a VTC network.