Research On Image Recognition Base On Neural Networks And Model Combination

The Neural Networks is an important model of machine learning, it use a hierarchical structure which consist of many perceptrons to fit a complicated nonlinear classification surface. Deep learing combines unsupervised feature learning and neural networks. In order to get a better classification result, it uses the features extracted from unsupervised feature learing to initial the neural networks.This paper researches on methods that can improve the accuracy rate of neural networks further, they includes changing active function, adding noise or dimension reduction on training set, adjusting the configuration parameters of neural networks. All of these methods do adjustment on a single model and require mathmetical foundation and experience. Besides, due to the lack of guidance and many changeable parameters, adjusting parameters to improve the accuracy rate is often a time-comsuming process. From the thought of ensemble learning, we proposed a model combination method based on imporved Adaboost algorithm. This method based on Adaboost algorithm, adds an upper limit for the weight of misclassified sample, this can prevent the misclassified sample from having too big weight so that if the model has misclassified it again, then it won’t get a very low model weight. Meanwhile, during training each model, the method calculates the accuracy rate for each class. At last, the method will use both model weight and accuracy rate of each class to decide the final result.Through integrating the Convolution Neural Networks and Deep Neural Networks, the experiment results show that this method greatly improve the accuracy rate if both models have a medium accuracy rate. Comparing with those model combination methods such as majority voting, weighted voting and method based on Adaboost algorithm, this method has a better result. Thus this method can be used to improve the final result if we get two or more neural networks which have a medium accuracy rate. Through this method we can avoid the complicated adjusting-parameter process which in order to get a best single networks model.