Visual saliency-based vehicle manufacturer recognition using autoencoder pre-training deep neural networks

Abstract

Vehicle manufacturer recognition (VMR), consisting of vehicle logo detection (VLD) and vehicle logo recognition (VLR), is now a crucial part of intelligent transportation system (ITS). A novel VMR method combining visual saliency detection and autoencoder pre-training deep neural network (AP-DNN) is proposed in this paper. An automatic VLD method based on visual saliency detection is used to build a vehicle logo dataset. This dataset contains 10000 training samples and 1500 testing samples for ten types of vehicle manufacturers. In the experiment stage, using AP-DNN, a VLR rate of 99.20% with a training time of 40 min is obtained, which shows higher accuracy than scale-invariant feature transform (SIFT) or AdaBoost-based methods and less training time than methods using a convolutional neural network (CNN). Further, with 2000 vehicle images for ten different types of manufacturers, a VMR rate of 97.95% is obtained automatically and demonstrates the robustness and efficiency of our method.