Research On Recognition Of Vehicle Logo

Nowadays, the blooming economy results in a better daily life for citizens in China. Therefore, to purchase a more convenient and comfortable travelling experience, owning a private car is in their urgent demand, which leads to a shape increase of car ownership. So far, China has become the world’s largest car producer and owns the largest car consumption market, which makes the regulation for vehicle is increasingly crucial. As a result, the intelligent transportation system has been rapidly developed. Due to the reason that vehicle-logo is one of the most significant symbol for vehicles, vehicle-logo recognition plays an important role in distinguishing vehicles. However, the robustness of vehicle-logo location and recognition is not satisfied for the following reasons:the image acquisition device, the complicated natural scene, the low resolution, light, shadow and night. Hence, we aim to handle the difficulties listed above and to improve the performance of vehicle-logo recognition, which could enhance the practicability of vehicle-logo recognition.First of all, we propose a coarse-to-fine approach for locating vehicle logo by considering license plate according to the prior knowledge, mathematical morphology, and the projection technology in binary image.Most of the existing logo positioning methods, which relays on license plate location, follow the pipeline:firstly search and inspect the location of the license plate, then discover the position of the vehicle-logo. This approach is computational expensive, at the same time, it is very sensitive to strong light, shadow and so on. To overcome these drawbacks, based on the rough position of the vehicle-logo, median filter is employed to remove the noise, and we use Sobel operator to detect the edge of vehicle-logo, then mathematical morphology operation is used to make the logo edge expanded and connected to form a connected area, finally projection is used to refine the location of vehicle-logo. The experiment shows that the proposed method is robust to strong light, shadow, and other natural situations such as at night.Secondly, we propose an improved texture descriptor based on LBP (Local Binary Pattern).The original LBP feature considers the neighborhood eight pixels, however, due to the pixel values in a continuous area are generally stable, we only pick up three points in the neighborhood eight pixels, which forms a three binary number, hence the image is segmented into several6*6small patches, each patch generates a eight dimensional LBP feature, which finally resulting a288(6*6*8) dimensional LBP feature vector. The improved LBP feature vector is capable to encode texture information, additionally, it introduces the spatial information, which offers a better representation of the vehicle-logo.In the end, we study a template matching algorithm based on the measurement of Euclidean distance and histogram similarity.Given the288dimensional LBP feature vector, Euclidean distance and histogram matching card coefficient are used to calculate the similarity between the target and template vehicle-logo, then the vehicle-logo recognition and classification are achieved. We discover a more suitable template matching algorithm for our proposed LBP feature descriptor, which introduces the spatial information, by experimental comparison.