Detecting License Plate In Natural Scenes Based On Cascade Convolutional Networks

With developing of information technology, images acquire and storage technology also has been great progress in social life, all kind of industries have accumulated a great deal of license plate images, most of these licenses plate image are obtained in a natural scene image which have high quality, diversification and big data. Locate the license plate area in a short period of time quickly, is to enhance the utilization of these license plate images. However, the existing license plate license plate detection techniques to obtain images of a scene, there are stringent constraints, these constraints in a natural scene are difficult to satisfy, so the study of natural scenes license plate detection technology has important practical significance.After read a lot of history in the technical literature, To solve the problems of detecting license plate in natural scenes with scene complexity, data intensity, various illumination and angle of view,this method propose a robust and fast architecture for license plate detection in natural scenes. By introducing convolutional network as the basic learning unit and further cascading the basic ones into more complex architecture, this method tries to achieve competitive detection rate with improved efficiency for the natural scene plate detection problem. Empirical results show that the propose method can yield high detection rate with efficiency on a real-world dataset under reasonable false positive rate.This paper considers the license plate detection system to build a natural scenes using concatenated convolutional network. Convolutional neural network is an artificial neural network, has become a hot topic in the field of image recognition. It’s weights shared network architecture reduces the complexity of the network model, reducing the number of weights. Convolution network is to identify two-dimensional shapes and special design a multi-layer perceptron, this kind of network structure on translation, scaling, tilt or a total distortion of his highly invariant form. For single convolution network, reducing neuronal can improve the detection efficiency, but it also may reduce the ability to distinguish the convolution of the network, reducing the detection accuracy. Cascade multiple classifiers to balance the classification accuracy and the classification efficiency is a simple and effective method.