No-Reference Light Field Image Quality Assessment Using Four-Dimensional Sparse Transform

Multi-exposure fusion light field image quality assessment with motion region detection

“…Therefore, in the end, only f 2 (2) and the mean of f 5 are retained, and in the following experiments, they will still be represented as f 2 and f 5 . The proposed method is compared with 16 representative IQAs, including 2D-IQA methods(BRISQUE [6] , GWH-GLBP [7] , GM-LOG [8] , NIQE [9] ), MEF-IQA methods (MEF-SSIM [10] , MEF-SSIM d [11] ),TM-IQA methods (BTMQI [12] , BTMIQA [13] , ISTMQI [14] ) and LFIQA methods (BELIF [15] , NR-LFQA [16] , Tensor-NLFQ [17] , TSSV-LFIQA [18] , PVRIs LFIQA [19] , PM-NLFIQM [20] , 4D_DCT-LFIQA [21] ). In the experiments, MEF-LFIs from the MEFLFD were randomly partitioned into an 80% training set and a 20% test set.…”

“…Liu et al [20] constructed an LFIQA model based on pseudo-reference sub-aperture images (SAI) and micro-lens images. In addition, Xiang et al [21] further measured the quality of LFI by extracting perceptual features from the 4D discrete cosine transform domain. Although these LFIQA methods consider the spatial distortions and angular inconsistency of LFI from different perspectives, they do not sufficiently take into account the artifact distortions presented in MEF-LFIs.…”

Deeblif: Deep Blind Light Field Image Quality Assessment by Extracting Angular and Spatial Information

“…In this subsection, the proposed DeeBLiF is compared with thirteen representative IQA metrics, including 5 types: NR 2D-IQA (BRISQUE [18], GWH-GLBP [19], and NIQE [20]), NR 3D-IQA (SINQ [21]), NR Multi-view-IQA (MNSS [22] and Wang's [23]), FR LF-IQA (MDFM [4] and Min's [5]), and NR LF-IQA (BELIF [8], VBLFI [9], NR-LFQA [10], Tensor-NLFQ [11], and 4D-DCT-LFIQA [12]). Following the evaluation method in [12], K-fold cross-validation is employed for all the learning-based NR metrics. For FR metrics and NR multi-view metrics, their performance is reported using the same test set as in K-fold cross-validation.…”

“…For other metrics, we cite their performance from [12]. Table 1 summarizes the overall performance comparison of state-of-the-art objective IQA metrics on Win5-LID dataset, where bold values indicate the best performance.…”

“…By contrast, the NR LF-IQA metrics are more challenging because they do not rely on any original image information, but they are also more suitable for practical applications since the acquisition of reference LFI is always limited. Most state-of-the-art NR LF-IQA metrics (e.g., [8,9,10,11,12]) firstly extract some well-designed hand-crafted features and then apply Support Vector Regression (SVR) to establish a mapping between the extracted features and the quality of distorted LFI. For in- 2) No CNN model specifically designed for LF-IQA.…”