Image Classification Based On Broad Learning And Deep Ensemble

Image classification is one of the core research fields of computer vision.In the era of information technology,there are new challenges in learning,mastering and using information.Information processing and communication are becoming more and more complex,while generation and dissemination are extremely rapid.Consequently,the data of image grows in a geometric rate every day.The management of images with rich information should also adapt to new characteristics and patterns.On these grounds,how to make the computer manage the massive image information more efficiently and classify the image data more quickly and accurately is a valuable research topic.This work proposes the following models and methods to optimize the effect of image classification based on broad learning,deep learning and ensemble learning:(1)An image classification method based on random broad learning network is proposed.Firstly,the orthogonal matrix generated randomly used to extract the features of the input layer constitutes the feature node of multi-layer and multi-branch neural network.Then,the extracted features and expected outputs are fitted according to the theory of sparse automatic coding.The Alternating Direction Method of Multipliers is used to solve and optimize the network parameters of each node.Finally,the prediction results of each branch network and each layer network are weighted to obtain the final classification results.The proposed model has no depth structure and is easy to solve,so the training time is short while the accuracy of image classification is high.(2)An optimization deep residual network model for image classification is proposed.Firstly,Several Squeeze and Excitation blocks is added to the residual network in layers,and different weights are assigned to the feature maps of different layers to improve the feature representation ability of network.Then,the cross-entropy function is used as the loss term,and the cross-entropy is used to fit a uniform distribution as the regularization term to prevent over-fitting and improve the generalization performance of the whole network.The proposed optimization model not only makes better use of the different advantages of the feature maps in different layers,but also improves the generalization ability of the model.The accuracy of image classification has been effectively improved.(3)An image classification method based on feature ensemble optimization of deep residual network is proposed.Firstly,the features extracted from several layers of pre-trained deep network model are sparsely and automatically coded to obtain the primary encoder weights by the Alternating Direction Method of Multipliers.Then the output of the primary encoder is coded in the same way to get the weights of the secondary encoder.Finally,the predicted results of the secondary encoders are integrated by weighting method to get the classification labels.The proposed optimization method only solves the encoder parameters without training the deep network,which greatly improves the training speed.Through the integration of encoders,the network feature representation ability is improved so that the classification accuracy is improved.