车牌超分辨率重建与识别

为了从图片中快速准确地识别车牌，提出一种结合图像超分辨率技术的车牌识别方案。车牌图片具有明显的特定的模式特征，只是具体的字符编码不同。因此车牌图片非常适合做超分辨率重建。本文提出的系统主要由车牌检测定位、车牌超分辨率重建、字符分割、字符识别等模块组成。综合基于边缘、基于颜色和基于最大稳定极值区域三种车牌检测策略并采用并行编程方法来综合检测结果得到候选车牌。采用车牌图片正负样本来训练支持向量机分类器。得到分类器模型后对候选车牌判决得到真正的车牌。随后对真实车牌图片进行超分辨率重建。该部分主要由基于固定邻域回归的方法实现。这种方法综合了稀疏字典学习和领域嵌入的方法，比较好的兼顾了准确率和计算速度。运用OpenCV提供的图像处理库来对重建后的图片做字符分割。得到单独的字符图片后采用人工神经网络进行识别。识别前先使用一定数量的字符图片对网络进行有监督训练获取识别模型。采用一个单隐层的神经网络，运用反向传播算法进行训练得到识别模型。最后提取字符图片的特征并输入网络进行分类完成识别。为了测试系统的表现，在实际场景中采集了一百张车牌图片作为测试集。实验表明，该系统具有较高识别准确率和较快的处理速度。

License Plate Super-Resolution Reconstruction and Recognition

On purpose of identifying the license plate quickly and accurately from an well captured image, a license plate recognition scheme with the image super-resolution technology is proposed. License plate pictures have similar pattern features though coded with different numbers and characters. Therefore, license plate images are very suitable for super-resolution reconstruction. The system proposed in this paper is mainly composed of license plate detection and location, super-resolution reconstruction, character segmentation, character recognition and other modules. The three license plate detection strategies based on edge detection, color processing and maximum stability and extreme region algorithm are synthesized with parallel programming skills to get the candidate license plates. The support vector machine classifier is trained by using positive and negative samples of license plate images in advance. After the classifier is obtained, the real license plate is selected with the prediction model. The real license plate is then reconstructed with super-resolution technic. This stage is implemented mainly by the method based on the anchored neighborhood regression. This method combines the advantages of sparse dictionary learning and neighborhood embedding. Thus the accuracy and speed of calculation are both well taken into account. The OpenCV library is employed in the project to do character segmentation for the reconstructed image. An artificial neural network is then employed on the recognition stage. Before recognition, a certain number of positive and negative samples of character images are prepared to train the recognition models. In this paper, we use two single hidden layer neural networks and train with back propagation algorithm. After the network is fine tuned the features of the numbers and characters from test images are sent to the networks to complete the finial recognition. In order to test the performance of the system, one hundred pictures of the license plate collected from the actual scenes serve as the test set. Experiments show that the system has high recognition accuracy and fast processing speed.