License Plate Recognition Super-resolution Algorithm

Because of photographs natural condition and factor imaging system's own limitation influences and other factors, making the license plate image occurs lower quality to some extent, thus affecting the rate of license plate recognition system. This paper using super-resolution image sequence reconstruction to improve the resolution of the plate image,super-resolution reconstruction of image sequence, is use of videos complementary information between adjacent frames, recovery the sequence image the detail information which loses in gathering and the transmission process, to improve video image resolution. Super-resolution image reconstruction generally to be composed image registration and image reconstruction of two parts.this article unifies the car license image the characteristic, uses carries on the image matching based on the SIFT special matching method. The SIFT algorithm to when carries on the image matching, must to extract the large scale space feature points and the calculated, so in the process of registration will appear the phenomenon of mismatch. Euclidean distance is usually the method used to eliminate the mismatch, through experiments found that, this method can not be good to eliminate false matching points when frames have large relative motion, to solve with this problem, this paper introduces the RANSAC algorithm on this basis, this approach eliminates the mismatch of feature points ability are relatively strong, But also has strong robustness.In this paper, the ministry of image reconstruction is used method of projection onto convex sets, this approach has a prominent feature is the priori knowledge contains capacity particularly strong, according to a priori knowledge contained in the image, establish a series of corresponding bounded convex set, constraint of the reconstructed image features such as smooth and reliable data. Through to comparison the effects of several commonly algorithm for reconstruction, POCS algorithm can preserve edges and details information well, meet the requirements of license plate image reconstruction. However, the noise reduction capability and stability convergence of the algorithm is not good, lead to the edges of the image appear ringing phenomenon after reconstruction, in this paper, add the regularization method to improve noise reduction capability in the standard POCS algorithm, introduction gradient of weight, use gradient descent method to minimize the relative error after a reasonable number of iterations, it was found through a large number of experiments, after adding regularization POCS algorithm can effectively reduce the edge ringing phenomenon, improve the effect of the reconstruction algorithm. Finally, introduces the various modules of license plate recognition system, implements a simple license plate recognition system, this paper using the method of BP neural network to recognize license plate character. Enter the image of before and after reconstruction to recognition system, used this paper raised super-resolution reconstruction method to improve the resolution of the plate image, to a certain extent to improve the accuracy of license plate recognition system.