Cold Start Problem of Vehicle Model Recognition under Cross-Scenario Based on Transfer Learning

Abstract

As a major function of smart transportation in smart cities, vehicle model recognition plays an important role in intelligent transportation. Due to the difference among different vehicle models recognition datasets, the accuracy of network model training in one scene will be greatly reduced in another one. However, if you don’t have a lot of vehicle model datasets for the current scene, you cannot properly train a model. To address this problem, we study the problem of cold start of vehicle model recognition under cross-scenario. Under the condition of small amount of datasets, combined with the method of transfer learning, load the DAN (Deep Adaptation Networks) and JAN (Joint Adaptation Networks) domain adaptation modules into the convolutional neural network AlexNet and ResNet, and get four models: AlexNet-JAN, AlexNet-DAN, ResNet-JAN, and ResNet-DAN which can achieve a higher accuracy at the beginning. Through experiments, transfer the vehicle model recognition from the network image dataset (source domain) to the surveillance-nature dataset (target domain), both Top-1 and Top-5 accuracy have been improved by at least 20%.