Infrared Image Super-Resolution Quality Prediction Driven by Frequency-Domain Features

In order to predict infrared image super-resolution (SR) quality under conditions of no SR image generation and no high-resolution reference image, this article proposes a fully connected neural network model for infrared image SR quality prediction driven by wavelet domain energy features. Utilizing the multiresolution and scale-invariant properties of undecimated wavelets, our model separates low-frequency information component and three high-frequency information components from a low-resolution image. Our model achieves decorrelation of image pixel information and generates wavelet domain energy normalization features. Utilizing a fully connected neural network, we construct a wavelet-based image SR quality prediction network. This neural network combines four subnetworks to enhance the network representation and learning capabilities. The network is trained using wavelet domain energy normalization features, while autonomously learning the mapping relationship between input data and evaluation metric. This model is validated on infrared image data sets. The experimental results indicate that the proposed model can accurately predict SR image quality metrics by utilizing frequency-domain energy features of a low-resolution image.