Wheel Model Identification Based On Compution Vision

Automobile wheel industry is an important part of China's auto manufacturing industry,it play an extremely important role in the development of the national economy.So this paper presents the wheel model identification algorithm based on computer vision,and we have studied the wheel model identification in many aspects,the main contents of the study include the following points.(1)Extracting the information of the number of wheel holes,radius of wheel,the area ratio of wheel holes to the whole wheel,and the wheels are classified by these information.(2)In the identification step of multi-feature fusion,the feature we extract is sample and the speed of identification performed well,but the identification rate is very low and far from the industrial requirements,so we proposed wheel model identification based on shape recognition and texture filtering,in the process of shape matching,we extract the shapes information of the wheels,first,set up a standard spoke shape template for each wheel,then the edge minimum distance is used to identify the same shapes of the identified wheels in the template library,and refer to the number of the wheel holes,if the number of shapes is consistent with the number of holes in the corresponding wheel,it is determined that the shapes are in accordance with the template wheel.(3)Random walk is used to verify of the matched wheels above,the method of random walk is used to describe the identified wheels and the corresponding wheels of the wheel library.Then we draw the random walk histogram of two pictures,calculating the variance of two histograms and put them in the same range.Finally,we determining whether the two wheels with the same wheel model by the deviation of two histograms.(4)However,the algorithm of wheel model identification based on shape recognition and texture filtering takes a lot of time,and it is difficult to reach the level of real time,so we proposed the algorithm of wheel model identification based on index weight VLAD features.First we extract the SIFT feature of the wheels,transforming the SIFT feature to RootSIFT,and then a large number of RootSIFT are clustered to form a code book,a VLAD vector will be formed if we Quantify all RootSIFT vectors in each wheel pictures to the code book,then we reduce the dimensions of VLAD by PCA,in this paper,we add an exponential weight for the drop-dimensional VLAD vector in order to reduce individual unstable value,finally,the most similar wheel is found by comparing the weighted VLAD vectors.