Infrared Image Complexity Metric for Automatic Target Recognition Based on Neural Network and Traditional Approach Fusion
Xiaotian Wang,Kai Zhang,Jie Yan,Muzeng Xing,Dongsheng Yang
DOI: https://doi.org/10.1007/s13369-020-04351-7
IF: 2.807
2020-01-24
Arabian Journal for Science and Engineering
Abstract:Infrared image complexity metric plays an important role in automatic target recognition (ATR) performance evaluation. In particular, with the development of the infrared imaging technology, there are many excellent infrared image complexity metrics for ATR. However, in the related works, there are two aspects of imperfections: (1) only the influence of individual feature is considered, ignoring the interaction among characteristics; and (2) these metrics all do not take the degradation of thermal imaging process into account. To overcome the imperfections, a novel criterion of evaluating infrared image complexity which considers the interaction among characteristics and the degradation influence is proposed. Firstly, to achieve complementary advantages among characteristics, the feature space is introduced to establish three image complexity indicators, respectively, namely feature space degradation complexity (FSDC), feature space similarity degree of global background and feature space occultation degree of local background. Each indicator is integrated by feature space to obtain complementary advantages. Secondly, to take the degradation of thermal imaging process into account, the neural network is trained to obtain the FSDC. In addition, the feature spaces are perfected by Pearson's correlation analysis and relevant features were removed so that each indicator is more reasonable. Finally, we connect the three image complexity indicators by using an improved analytic hierarchy process. The experimental results show that the proposed algorithm is more consistent with the actual situation than traditional statistical variance and signal-to-noise ratio.
multidisciplinary sciences