Infrared ultraspectral signature classification based on a restricted Boltzmann machine with sparse and prior constraints

摘要: The state-of-the-art ultraspectral technology brings a new hope for the high precision applications due to its spectral resolution. However, it comes with challenges brought by improvement of resolution such as Hughes phenomenon and over-fitting issue, our work is aimed at addressing these problems. As Markov random field MRF models, restricted Boltzmann machines RBMs have been used generative models many different pattern recognition artificial intelligence showing promising outstanding performance. In this article, we propose method infrared signature classification based on RBMs, which adopt regularization-based techniques improve accuracy robustness noise compared traditional RBMs. First, add an arctan-like term objective function sparse constraint accuracy. Second, utilize Gaussian prior avoid problem. Third, further performance, multi-layer RBM model, deep belief network DBN, adopted classification. Experiments using libraries provided Advanced Spaceborne Thermal Emission Reflection Radiometer ASTER Environmental Protection Agency EPA were performed evaluate performance proposed comparing other methods, including coding-based classifiers binary coding BC, feature-based SFBC, derivative feature SDFC matching novel extraction termed crosscut CF, three machine learning methods deoxyribonucleic acid DNA-based ADSM, SparseDBN. Experimental results demonstrate that superior was can simultaneously