Study On Band Selection Algorithm Of Human Hyperspectral Image In Pressure State

The study of pressure identification for non-contact physiological signals is an important part of affective computing.And the tissue oxygen saturation extracted by hyperspectral technique can be used for non-contact pressure identification.Hyperspectral image have a lot of information,a large amount of data and high similar band correlation.Since these shortcomings has brought great difficulties to the real-time pressure identification,reducing the hyperspectral data become the key.Band selection is a commonly used method for the band dimensionality reduction.This paper is study on the band selection algorithm for the human facial hyperspectral image.Details are as follows:(1)A hyperspectral band selection algorithm based on Physarum network is proposed.Physarum can solve the shortest path problem,and which is similar to the search for minimum redundancy bands in the hyperspectral band selection.In this paper,the Physarum algorithm is applied to the hyperspectral band selection.In order to realize the automatic band selection,a non-uniform subspace method which based on mutual information was proposed.Besides,the inverse of optimal index factor(IOIF)is used as adaptive function in this band selection method.The experimental results show that the non-uniform subspace method is better than the homogenization method,and the bands selected by IOIF adaptive function is better than that of the correlation.The identification of the bands selected by the Physarum algorithm is close to that of the whole band.The validity of the algorithm is proved.(2)A combination of linear prediction and tabu search for hyperspectral band selection is proposed.The bands obtained by linear prediction method are used as the initial solution of tabu search,and the classification accuracy of pressure is used as the fitness function.Then use the tabu search to obtain bands with a higher pressure classification accuracy.The experimental results show that bands selected by the tabu search algorithm are more accurate than those only selected by linear prediction.The effectiveness of the algorithm is proved.(3)Design an experiment for induced psychological stress,and obtained 25 ordinary college students facial hyperspectral data which including psychological stress state and calm state.A method for automatically collecting spectral curves based on oxygen histogram was put forwarded.Besides,the minimum noise separation method is used to remove the noise of the obtained hyperspectral data.The experimental results show that the tissue oxygen saturation map generated by denoised data is better than no-denoising data,and it is more beneficial to identify the pressure state after the band selection.In this paper,the collected hyperspectral data was used to verify the two band selection algorithms.It is shown that the two band selection algorithms proposed in this paper can maintain the accuracy of pressure identification and reduce a large number of bands.These two band selection algorithms are effective for pressure identification which based on hyperspectral technique.