Deep Convolutional Neural Network For Car Recognition

Security monitoring system is a security monitoring system in the particular place, which can record and process cars passing it. We call the front part of car "face" and the security monitoring systems mainly take face images of the cars. This system can retrieve the type of car and find out the owner of the car by the number of license plate. So security monitoring system has important research significance and application value in intelligent transportation and it is highly valuable for social security. The traditional methods in the car identification are manual search or car plate recognition technology. Manual search is tedious and time-consuming. Car plate recognition technology is difficult to occluded license plate or illegal license plate. Thus, it has great potential to study an automatic car recognition method to perform the task. This thesis is prepared to address this problem based on the car face images taken from the security monitoring system. These images can effectively represent attributes of the car. So use the car face to identify the type of car has become a hot research topic recently. However, there are also some challenges. The differences of car faces between different brands or different types are huge, and the differences of car face with same type and different configuration is small. Car faces in the image vary in color, position, and angle. These factors make car recognition become difficult. To solve these problems, we propose a car face recognition method based on convolutional neural network, which relies on the appearance of car face to identify the car’s type quickly and accurately. The main contributions of this paper are as follows:(1) We collect and label a car face image database, which contains a total of more than 70,000 car face images and 315 classes. Generally car face images are not public. And they are not easy to get. So this car face database has great research value.(2) We annotate 33 feature points for the car face including car face contour, windshield, rearview mirror and license plate. And we design a car alignment algorithm based on the constrained local model. This model can detect car face feature points quickly and accurately, and align all car face images to the same scale. The constrained local model consists of two sub-models:shape model and the feature model. The shape model describes the shape of the car face. The feature model describes the feature of the car face.(3) We design and implement a convolution neural network model for the car face images. This model uses full car face image as input to get car face feature. Based on this basic model, we study the effects of different components of the car face for the recognition rate, and propose a combination convolution neural network model based on the multi-component. This combination model divides car face into 5 components, and for each of these components we train convolutional neural network sub model separately. Finally, we get the best feature to describe the whole car face through a combination of sub-model. Then use the integrated features for car face identification and classification.We evaluate our method on the collected dataset. And the result shows that the car alignment algorithm based on the constrained local model can increase the rate of face recognition significantly. Compared with the unaligned car face data, the recognition rate of aligned car face data is increased by 8%. The car face recognition algorithm based on the combination convolution neural network model has a better description for car face than the traditional HOG feature and get a 92% car face recognition rate.