Non-reference quality assessment of infrared images reconstructed by compressive sensing

Infrared (IR) images are representations of the world and have natural features like images in the visible spectrum. As such, natural features from infrared images support image quality assessment (IQA).¹ In this work, we compare the quality of a set of indoor and outdoor IR images reconstructed from measurement functions formed by linear combination of their pixels. The reconstruction methods are: linear discrete cosine transform (DCT) acquisition, DCT augmented with total variation minimization, and compressive sensing scheme. Peak Signal to Noise Ratio (PSNR), three full-reference (FR), and four no-reference (NR) IQA measures compute the qualities of each reconstruction: multi-scale structural similarity (MSSIM), visual information fidelity (VIF), information fidelity criterion (IFC), sharpness identification based on local phase coherence (LPC-SI), blind/referenceless image spatial quality evaluator (BRISQUE), naturalness image quality evaluator (NIQE) and gradient singular value decomposition (GSVD), respectively. Each measure is compared to human scores that were obtained by differential mean opinion score (DMOS) test. We observe that GSVD has the highest correlation coefficients of all NR measures, but all FR have better performance. We use MSSIM to compare the reconstruction methods and we find that CS scheme produces a good-quality IR image, using only 30000 random sub-samples and 1000 DCT coefficients (2%). In contrast, linear DCT provides higher correlation coefficients than CS scheme by using all the pixels of the image and 31000 DCT (47%) coefficients.