Study On Densely Connected Deep Extreme Learning Machine Algorithm

The current deep Extreme Learning Machine(ELM)algorithm has the following disadvantages: when the number of network layers of the algorithm is shallow,the random feature mapping makes the sample features cannot be fully learned and utilized;when the number of network layers of the algorithm is deep,the validity of the sample features will decrease after continuous abstraction and generalization.Based on the existing deep ELM algorithm,this paper proposes two densely connected deep ELM algorithms: Dense-HELM algorithm and Dense-KELM algorithm.And the Dense-KELM algorithm proposed in the paper is used for remote sensing image scene classification.The basic structure of the Dense-HELM algorithm is an ELM-based automatic encoder(ELM-AE).The basic structure of the Dense-KELM algorithm is a K-ELM-based automatic encoder(KELM-AE),and the Dense-KELM algorithm is improved based on the Dense-HELM algorithm.The experimental results show that the densely connected network structure allows the algorithm to fully extract the feature information of the sample when the number of network layers is very shallow,which significantly improves the recognition accuracy of the algorithm.And as the algorithm improves the utilization of sample features during training,the number of hidden layer neurons can be reduced in an appropriate amount,which further accelerates the training speed of the algorithm.The main work of the thesis is divided into the following parts:The first part briefly introduces the research background and significance of the thesis,the statuation of research and development in the domestic and overseas,as well as the main research contents and chapter structure of the thesis.The second part introduces the theoretical knowledge related to the subject research.The third part introduces in detail the mathematical theory,model structure and training process of the Dense-HELM algorithm proposed in the paper,designs the comparison experiments of the algorithm,and analyzes the experimental results.The fourth part introduces in detail the mathematical principle,model structure and training process of the K-ELM-based Dense-KELM algorithm proposed in the paper,and the performance comparison between Dense-KELM algorithm and other deep learning algorithms.The fifth part introduces the practical application of remote sensing image scene classification based on Dense-KELM algorithm.The results are compared with other deep learning algorithms to verify the practicability and effectiveness of the proposed algorithm.The sixth part sums up the whole work of the paper and prospects the future follow-up research directions.