In order to solve the problem that LNSCT loses the target contour information due to discarding the low frequency components of image

a new illumination invariant extraction method

called MLNSCT

is proposed. Firstly

NSCT is used to decompose the input image in logarithm domain to realize the separation of the low-frequency and high-frequency components. Secondly

the BayesShrink threshold filter is applied to the high-frequency sub-band coefficients

and the inverse NSCT is performed for the low frequency components to obtain the feature image. Thirdly

a second NSCT decomposition on the feature image

the threshold filtering on high-frequency sub-band and inverse NSCT on low-frequency component are performed sequentially. After multiple NSCT decomposition

the illumination invariant features of the input image are extracted from the set of all high frequency sub-band coefficients. Through further study of the relationship between the illumination invariant features and the raw image

Dual Lenet

which is a parallel synchronous convolutional neural network

is designed to improve the accuracy of ground target recognition by fusing the high-level features of both. The experimental results show that MLNSCT has higher classification accuracy than that of LNSCT in Lenet model

and the classification accuracy is higher with the increase of decomposition number. Furthermore

it is proved that the fusion of illumination invariant features and raw image can effectively improve the classification accuracy of ground target recognition.