Recognition Of Vehicle-logo Based On Intelligent Image Processing

The vehicle-logo is an important auxiliary characteristic for vehicle recognition,with the advantages of strong identifiability,distinctive features and hard to change.Consequently,the further study for vehicle-logo recognition,which is one of the key technologies of intelligent transportation system,will be valuable to theory and application.Vehicle-logo recognition system consists of two parts:vehicle-logo location and vehicle-logo recognition.Vehicle-logo location is mainly based on prior knowledge,vehicle-logo locating methods only depending on the relative position between vehicle plate and logo have many drawbacks;the classification method of vehicle-logo recognition is mainly based on support vector machine(SVM)and convolutional neural network(CNN),which cannot meet the increasing recognition accuracy and real-time demands.For those problems mentioned above,a vehicle-logo location method based on visual attention mechanism combined with improved ant colony algorithm and vehicle recognition method based on hybrid CNN model are proposed to achieve the purpose of accurate location and recognition of vehicle-logo in this paper.The main contents are as below:Firstly,a vehicle-logo location approach based on visual attention mechanism combined with improved ant colony algorithm is proposed.And the optimization design of the model is developed in the process of saliency calculation and focus transferring path.The multi-scale residual spectrum method is employed to calculate the global salient region,and then three new indicators including quality symmetry degree,the compositional complexity of image and the shape complexity are used to measure the complexity of sub-area,therefore the degree of differentiation between vehicle-logo and the background is observably increased;focus of attention are generated after competition of salient sub-regions,then sub-area complexity and ant false-rate are introduced to drive ants to visit the point with bigger area complexity preferentially in order to optimize the focus transferring path,accelerate the convergence speed and alleviate the impact of falling into local optimum easily.Finally,the area of vehicle logo is judged and segmented,according to the calculated optimal path.Secondly,vehicle-logo recognition method based on hybrid CNN model is proposed to improve the recognition accuracy and the performance of real-time.The method of inter-class variance of complexity are utilized to classify the vehicle-logo samples before the input layer,the model needs to train t-wo traditional convolutional neural networks at the same time,sigmoid function is adopted by the one with small complexity difference and tanh fonction is selected by the other one with large complexity difference.And cross entropy is selected as the network cost function to obtain a faster tap-weight updating speed.The vehicle-logo feature vectors learned from the two traditional CNN are coalesced in the whole connection layer for final classification and recognition.The simulation experiment results show that the vehicle-logo location rate of the proposed method reached 98.43%.The method has good performance of robustness and real-time,which is able to satisfy practical requirements.Simultaneously,the vehicle-logo recognition rate of hybrid CNN model reached 98.61%,and it also has a satisfactory recognition effect for vehicle-logo in condition of rotation and translation,Gaussian noise and uneven brightness,with a good anti-noise performance.