A Study on the Recognition and Classification Method of High Resolution Remote Sensing Image Based on Deep Belief Network

Abstract

High resolution remote sensing images can describe the geometric features, spatial features and texture features of objects more accurately, which are widely used in various fields. How to get more useful information from the remote sensing image, and then the recognition and classification of the image from the information has become one of the hot spots in the field of high resolution remote sensing image research. Deep learning is a learning algorithm based on the depth network structure, which can better fit the intrinsic structure of the sample, compared with the traditional shallow classifier. Depth of learning in a deep belief network model is based on single-layer Boltzmann machine learning algorithm, each layer is made up of the generation and cognition, and make the bidirectional weight updatin g come true, the net output of each layer can be reduced to the input signal, so that the model can be infinitely close to the global optimum in the pre training stage. The author propose an improved dropout strategy based on the study of deep belief network model, this strategy only chooses partial local area data to zero out the weight at each time. It not only maintains the local information of the image itself, but also enhances the generalization ability of the model. The experimental results show that the improved dropout strategy improves about 2.5% of the classification accuracy, and it has better classification performance.