Abstract: LPR(License Plate Recognition) in nighttime is one of the challenging works for developing artificial vehicle systems. Traditional license plate recognition algorithms have a low accuracy due to uneven lighting and camera shake exposure at nighttime. This paper proposes a novel license plate recognition algorithm based on improving light streak deblurring method, which introduces a constraint on the linear fuzzy model to estimate the fuzzy kernel. The method is able to estimate the blur kernel and deblurred image more accurately, and output image computed by the improved algorithm has better contrast ratio and resolution. We conduct the experiments on the benchmark license plate database, experimental results show that the proposed method in this paper is able to recognize license plates in low-illumination environments which has better resolution and contrast ratio with little interference resulted by noise, besides that its average kernel similarity and peak signal-to-noise ratio are significantly better than state-of-the-art works. The recognition accuracy improved at least 22. 6% compared the result without de-blurring.

Key words: license plate recognition')">

license plate recognition, light streak de-paste, convolutional neural networks