Learning Based Vehicle License Plate Image Super Resolution Reconstruction

Nowadays, surveillance systems are widely used in cities to monitor traffic situations and record accidents. In these applications, research about vehicle license plate is always of great interests to people for it can uniquely identify a car. However, due to both the far distance between camera and cars and blurry resulted from motion, the obtained car plate images tend to be of low resolution and subsequently hard to identify. As a technique used to improve the quality of low resolution image, super resolution can solve the foregoing problem.In this paper, based on the drawbacks of single frame sparse representation, a novel multiple frames joint learning strategy is proposed, which can reconstruct better results. Then, we put our attention on car plate super resolution. A synthetic vehicle plate image generation software is developed. The synthetic car plate sequence cannot only be used to serve as sample images in sparse representation, but can also be used in a dynamic PCA strategy as training segments. Different from traditional PCA, the dynamic PCA can adaptively select out optimal representation bases.First, a synthetic car plate image generation software is developed. It can transform the image and add noise, blurry to the generated sequence. Using the synthetic sequence cannot only improve the performance of the sparse representation algorithm, but can also help us to make it clear the relationship between selection of sample set and the performance of the algorithm.Beyond that, a multiple frame joint learning strategy is proposed based on the draws of single frame learning. Based on the assumption that low resolution patches from the same high resolution patch, between which there are subpixel translation, can hold similar structure. Taking advantage of some well-matched patches together to recover the sparse coefficients can avoid over-learning and is more robust to noise.Finally, a dynamic PCA based super resolution algorithm is presented. It attains great reconstruction result through adaptively narrowing the training samples.