Deep Network Model Based On Random Forest And Its Application To Image Classification

Image classification has always been a challenging area of research.The most essential part of it is feature extraction.It is very important to find efficient feature extraction algorithms in the image field.The ability of generalization of features previously selected manually is limited,and new features are often needed to deal with new problems.Deep learning have been widely used since 2006,it learn from the layered information processing mechanism of human brain vision,and automatically extracts the low-level to high-level features of images through multilayers neural networks.So it became a good way to overcome manually select features.For the deep learning itself,there are problems such as complicated structure,difficult operation,and long training time.This paper has done the following research work:Firstly,this paper studies deep network architectures with random weights to solve the problem of long training time in deep learning,and we propose a deep model with random weights based on Random Forsets(RF).RF is a combination classifier,which solves the problem of overfitting in the decision tree,and it runs faster and can maintain high efficiency for a large number of data.In the model C-RF,the convolution layer and the pooling layer with random weights are used to extract the features of the image,and then the classification is made in the RF.The experiment shows that the algorithm not only greatly shortens the training time of deep learning,but also ensures the classification accuracy.Secondly,the existing convolution neural network structure is complex and not easy to operate,it also requires the large amount of training process and parameter setting experience to operate.So an image classification method based on kernel principal component analysis network(KPCANet)is proposed in this paper.In the new model,the KPCA filter is used as the basic convolution filter,the nonlinear layer is set to be the simplest binary quantization(hashing);for the feature pooling layer,we simply use the block-wise histograms of the binary codes,which is considered as the final output features of the network,and the classification is done by Random ELM Forets.Experiments show that our model further improves the performance of the convolutional neural network and reduces the complexity of setting artificial parameters.