License Plate Detection Algorithm Based on Nearest Neighbor Chains
基于最近邻链的车牌检测算法

This paper proposes a nearest neighbor chain based method for adaptive license plate (LP) detection according to the geometrical characteristics and spatial arrangement of license plate characters. Firstly, image is segmented with adaptive binarization method to avoid the problem created by nonuniform illumination, and some undesired image areas are removed by limiting the range of region properties of connected components. Secondly, nearest neighbor pairs are constructed according to region properties of connected components of LP characters, and they are merged into nearest neighbor chains, thus, to detect all the candidate license plate regions. Finally, two variable-length square templates are designed to match horizontal and vertical projections of LP image gradients. It can verify all the candidate LP regions and determine the boundaries of all LP characters simultaneously. Experiments show that this method can deal with nonuniform illumination, LP size change, rotation, background interference, quality degradation and so on, and that it has achieved desired detection result for various types of LP images in complex scenes.