A Category-consistent Deep Network Learning For Accurate Vehicle Logo Recognition

Vehicle logos are important vehicle manufacturer information,which is hard to be replaced.These can be used as significant vehicle identification features and play a crucial role in intelligent transportation systems.The detection and identification of vehicle logos are of great significance to solve crimes and vehicle management.At present,a large number of detection and recognition methods have been proposed,and these methods have been applied in some traffic scenes.However,due to the variety of vehicle logos and the complexity of imaging conditions,these methods mentioned above are not robust and accurate enough in various conditions.This paper proposes a category-consistent deep network learning framework for accurate vehicle logo recognition,which can achieve high accuracy in multi-category and various situations.Compared with existing methods,this paper mainly has three contributions.First of all,this paper proposes a vehicle logo feature extraction neural network.The network takes advantages of the identity shortcut connection and the dense connection,making full use of low-level and high-level features.This network efficiently extracts features from shallow and deep layers.In training,it can alleviate model degradation and help gradients flow easily to promote the network learning and parameters update and improve vehicle logo recognition performance.Secondly,a category-consistent mask learning module is proposed.This module can predict binary masks of category-consistent regions corresponding to input images.These generated masks contain common visual features under the same category of vehicle logos.By enforcing the network to predict category-consistent regions,this module can help the network to learn discriminative category features.Thirdly,this paper proposes a category-consistent deep network learning framework by combining the above two modules and a classification module.In this framework,the network is enforced to predict category-consistent regions when training for recognition.This learning strategy can help the network pay more attention to the discriminative features of each vehicle logo category to strengthen the discriminative features learning.This paper conducts a large number of comparisons on existing approaches in five public vehicle logo datasets(HFUT-VL1,HFUT-VL2,XMU,Comp Cars,and VLD-45)to verify the feasibility and superiority of the proposed method.By employing the proposed VLF-net and learning framework,our algorithm achieves accuracies of 99.56%,98.73%,100.00%,99.92%,and 92.63% for the five datasets,respectively.A number of experiments and corresponding analyses demonstrate that the proposed deep network learning framework can improve the recognition accuracy of deep learning-based methods in vehicle logo recognition.It can enhance the discriminative features learning of the VLF-net and be applied to most of the existing popular networks.The proposed unified algorithm can achieve better performances for recognizing vehicle logos from vehicle logo images and frontal images of vehicles,which has research values and significance.