Research Of Vehicle Logo Recognition Technology Based On Convolution Neural Network And Processing Strategy On Few Samples

As the number of domestic automobile increases, establishing an efficient traffic management mechanism is imminent and the vehicle recognition technology arises at the historic moment. Vehicle logo is an important feature which contains the information of vehicle manufacturers, and for the difficulty of being replaced, it plays an auxiliary role in the vehicle recognition.Present vehicle logo recognition algorithms are mostly shallow structure, their expression ability of complex function is limited for the calculation cell is limited. Deep learning technology can achieve the complex function approximation through learning a nonlinear network structure which shows a strong ability to learn essential characters of data. Convolution Neural Network as a typical model of deep learning network, combines the feature extraction and classification process. It can find the local characters of data and maintain the invariance of translation, rotation and scaling by its receptive field, weight sharing and sampling strategy. When the training samples are enough, convolution neural network has an excellent performance in image recognition task. But when the training data is few, it shows a strong dependence on large samples. Therefore, studying how to achieve better performance on few samples has important value.Basing on basic concepts of Artificial Neural Network and Convolution Neural Network, a typical convolution neural network model is applied to vehicle logo recognition. Comparing with the traditional vehicle logo recognition technology which features are extracted manually, the scheme proposed in this work can extract features independently, and the image can be input directly. In allusion to the character that the Convolution Neural Network depends on the large samples, this work puts forward three kinds of strategies on few samples. The main contents and innovations of this thesis are as follows:1) Two kinds of schemes based on the disturbance to extend the training samples are proposed. One is adding noise and doing geometric transformation on the few samples to produce new samples. The other is adding disturbance in the feature subspace of every one kind of samples, reconstructing the new samples in the disturbed feature subspace.2) According to the Convolution Neural Network’s character of hierarchic structure and receptive field. This work examines the network’s performance of different convolution kernel sizes and the different scales of Convolution Neural network on the few training samples.3) A transfer learning strategy is proposed. Different image recognition tasks and unrelated data sets often have commonly used visual patterns. A good initial state can improve the performance of Convolution Neural Network. In this thesis, the MNIST handwritten digits is used to train a Convolution Neural Network model, and use the model to train the vehicle logo.The experiment proves that the Convolution Neural Network can be well applied to the vehicle-logo recognition system. The structure of the network has great influence to classification performance when the training sample is few. Sample expansion and transfer learning can improve the classification performance under the circumstances of small training samples.