KonIQ-10k: An Ecologically Valid Database for Deep Learning of Blind Image Quality Assessment

Visual saliency computation and quality evaluation via deep learning

“…This sub-task intends to make use of the distortion category information which is available in the legacy IQA datasets with some common synthetic distortions [11,161,162]. However, the massive Internet images captured by real cameras are usually afflicted by complex mixtures of multiple authentic distortions [7,163], which cannot be well-simulated by the limited algorithm-generated distortions in these legacy IQA datasets. As a result, such a distortion identification sub-task cannot accurately identify the complex mixtures of distortions existing in authentically distorted images and may lead to performance degradation when applying to evaluate the real-world images with diverse authentic distortions.…”

“…This inspires us to incorporate the visual saliency prediction as the auxiliary sub-task to learn a powerful multi-task deep IQA model for the quality evaluation on the authentically distorted images. [7], and larger MOS (mean opinion score) shown in the bottom indicates better subjective perceptual quality. Their saliency maps in the second row are generated by our DINet [8] and fused with the original images where a pixel with brighter intensity indicates a higher probability of attracting human visual attention.…”

“…bined with our PSG loss. For each loss coupled with our PSG loss, the improved or degraded results are marked in red or green, [7] . .…”

“…. 96 5.7 PLCC results of saliency models on KonIQ-10k IQA dataset [7] with our ResNet-based SGDNet and MIT300 saliency dataset [12] Visual attention mechanism refers to the ability of the human visual system (HVS) to automatically select the most salient or informative regions from natural scenes by filtering out redundant and unimportant visual information for further processing. Around 10 8 -10 9 bits per second of visual data enters into our eyes, as reported in [13].…”

Adaptive fusion of bright-field microscopy images acquired in different focal planes

“…Benchmark datasets are fundamental in computer vision and image processing research in order to track the performance, accuracy and efficiency of new methods and algorithms. The image quality assessment was also evaluated against literature methods with the well-known Computational and Subjective Image Quality database (CSIQ) image quality assessment database (LARSON; CHANDLER, 2010) and the KonIQ database, the largest image quality assessment database to date (Hosu et al, 2020).…”