Hyperspectral image quality evaluation using generalized regression neural network

Abstract In order to alleviate the overfitting problem caused by image quality evaluation (IQA) model learning under intolerably small data set, this paper proposes a multi-feature fusion-based deep architecture for hyperspectral image quality assessment. First, eight key IQA-related features, which are descriptive to the mean noise of multi-band images, spatial correlation, inter-spectral correlation, blur, and the phase-consistent map of images, are extracted from each hyperspectral image within the dataset. Based on this, a carefully-designed generalized regression neural network (GRNN) with a limited number of parameters is hierarchically trained by the feature vectors from samples in the training IQA data set. Comprehensive experimental evaluations on the hyperspectral IQA images from the DOTA dataset and the EO-1 Hyperion dataset have shown that the proposed model can indicate the subjective/objective quality-aware images regions precisely In addition, we observe that our designed IQA method has received impressive IQA performance than the other state-of-the-art non-reference methods.