The framework for this research work is the acquisition of Infrared (IR) images from Unmanned Aerial Vehicles (UAV). In this paper we consider the No-Reference (NR) prediction of Full Reference Quality Metrics for Infrared (IR) video sequences which are compressed and thus distorted by an H.264 codec. The proposed method works as a Bitstream Based (BB) approach and it may thus be applied on-ground. Three different types of features are first computed: codec features (based on information extracted from the bitstream), image quality features (based on BRISQUE evaluations) and Spatial and temporal perceptual information. Those features are then mapped, using a machine learning (ML) algorithm, the Support Vector Regression (SVR), to the quality scores of Full Reference (FR) quality metrics. The novelty of this work is to design a NR framework for the prediction of quality metrics by applying ML algorithm in the IR domain. A set of 5 drone energy leakage image sequences and 3 ground IR image sequences are used for evaluating the performance of the proposed method. Each of the image sequences are encoded at 4 different bitrates and the prediction of the proposed method is compared with the true FR quality metrics scores of four images metrics: PSNR, NQM, SSIM and UQI and one video metric: VQM. Results show that our technique achieves a fairly reasonable performance. The improved performance obtained in SROCC and LCC is up to 0.99 and the RMSE is reduced to as little as 0.01 between the actual FR and the estimated quality scores for the H.264 coded IR sequences.
Kabir Hossain, Claire Mantel, Søren Forchhammer, "No Reference Prediction of Quality Metrics for H.264 Compressed Infrared Image Sequences for UAV Applications" in Proc. IS&T Int’l. Symp. on Electronic Imaging: Image Quality and System Performance XV, 2018, pp 108-1 - 108-6, https://doi.org/10.2352/ISSN.2470-1173.2018.12.IQSP-108