Vehicle Image Classification And Recognition Based On Sparse Coding

As an important aspect of the intelligent transportation system, vehicle classification and recognition technology has a wide application in practice. Traditional methods of vehicle classification and recognition based on geomagnetic sensor is measured on the models directly, access to some of the physical properties of the models. The disadvantage of this approach is a large number of auxiliary measurement equipment need to be set up. In this thesis, the method based on vehicle image to achieve recognition and classification of the models, compared to the traditional methods, this method is easier to operate by using image. The sparse coding methods take advantage of the sparsity of the image, encode global features or local features of the image, sparse coding with a smaller quantization error. The adaptive coding means on the image itself or the features of the image can capture more prominent information.How sparse coding of images? It has become one of the challenging topics of the field of image classification. This thesis studies the vehicle classification recognition based on the model image sparse coding. The specific research works are as follows:Firstly, it is presented that a model of image classification method which uses the global feature of image sparse representation. The image of the sample set in accordance with each category to achieve a sparse linear combination with the use of global features to sparse representation. This method uses the nearest neighbor classifier. It is relatively simple and easy to implement and has a fast identification speed.Secondly, on the basis of the traditional coding of the Bag of Words model, this thesis proposes a method which based on sparse coding representation of the models local image feature descriptor. By means of coding and pooling local features, it can get a global histogram. By using support vector machine as classifier, the classification and recognition accuracy rate of this method is high, but the computation is large.Finally, sometimes the difference between the images is relatively small, this thesis introduces the idea of hierarchical image classification and design a multi-class classifier. First of all, confusing samples classified as a class, and then meticulous division. Using the methods of multi-class classifier combined to divid image hierarchically, it can further improve the average of classification accuracy.