Vehicle Recognition Method Based On Machine Learning

Vehicle target recognition has potential application in civil field as well as military field. However, for the resolution limit of source images, for example, satellite images, remote sensing images and so on, there are few detail-information of vehicles in them. Vehicle cannot be detected or recognized by using model methods.A vehicle recognition method based on machine learning, which is totally based on gray characters of pixels of objects and their surroundings, and does not need model characters, is proposed in this thesis. Firstly, target regions are separated from backgrounds by using image segmentation algorithms, and regions are searched by using seed-fill algorithms. Then the regions which have been searched through experiential rules are filtered, and region data is changed to vector data. Subsequently, these vector data are input into classifiers, including back propagation neural networks and support vector machine, which have been trained, to recognize. At last, the recognition results are signed in image and information of targets are collected.Experiments show that this method based on machine learning achieves high recognition rate and low false-rate and missing-rate. And it can run efficiently, satisfies the real-time request and has application values.