--- title: CID2013 in Image Quality Assessment url: https://www.emergentmind.com/topics/cid2013 type: topic --- # CID2013 in Image Quality Assessment CID2013 is a benchmark identifier that appears in recent image quality assessment literature as a human-rated image corpus for no-reference prediction, subjective score preprocessing, and perceptual quality regression. In the works considered here, CID2013 is used in three technically distinct settings: blind image-sharpness assessment with Kolmogorov–Arnold networks (KANs), perceptual-similarity-based preprocessing of subjective scores, and no-reference image contrast assessment with customized convolutional backbones. The later literature does not present a single uniform description of CID2013; rather, each study specifies its own image count, distortion taxonomy, and scoring protocol, while consistently treating the dataset as a mean-opinion-score-supervised benchmark for quantitative evaluation [2409.07762], [2404.19666], [2509.21967]. ## 1. Reported corpus definitions in later literature Across the cited works, CID2013 is described in three different ways. In the KAN-based blind image-sharpness assessment study, it is a database of 474 distorted photographs with realistic distortions collected from consumer devices, including motion or defocus blur, compression, sensor noise, and processing artifacts, with MOS normalized to the range \([0,100]\) [2409.07762]. In the perceptual-similarity score-preprocessing study, it is reported as containing 981 distorted images derived from 24 pristine originals, with five standard distortion classes: global contrast change, additive white noise, JPEG compression, lens blur, and color quantization with dither [2404.19666]. In the contrast-assessment study, it is described as 400 consumer-camera photographs captured by 79 different devices and organized into six scene categories—portraits, landscapes, architecture, natural objects, textural patterns, and indoor scenes—with authentic contrast variations arising from exposure settings, device tone-mapping, sensor noise, and JPEG compression [2509.21967]. | Source | Reported CID2013 composition | Reported task context | |---|---|---| | [2409.07762] | 474 distorted photographs; realistic distortions from consumer devices | Blind image-sharpness assessment | | [2404.19666] | 981 distorted images from 24 pristine originals; five distortion classes | Subjective score preprocessing | | [2509.21967] | 400 consumer-camera photographs from 79 devices in six categories | No-reference contrast assessment | This divergence is methodologically important because image count, distortion structure, and scoring protocol directly affect model capacity, train-validation-test design, and metric comparability. A plausible implication is that reproducibility on CID2013 requires paper-specific verification of the exact corpus instantiation rather than reliance on the acronym alone. ## 2. Subjective scoring protocols and target variables The sharpness-assessment study uses single-stimulus mean opinion scores normalized to \([0,100]\), obtained via crowdsourced pairwise or rating studies as described in Virtanen et al. (2014) IEEE T-IP [2409.07762]. Under that formulation, the regression target is a continuous MOS suitable for PLCC and SRCC optimization and evaluation. The paper’s feature matrices make this explicit: the mid-level matrix is $$ M = [\,x_{1,1}\ \ldots\ x_{1,15}\ |\ y_1;\ \ldots;\ x_{n,1}\ \ldots\ x_{n,15}\ |\ y_n\,], $$ and the high-level matrix is $$ N = [\,f(I_1)\ |\ y_1;\ \ldots;\ f(I_n)\ |\ y_n\,]. $$ The perceptual-similarity preprocessing paper treats CID2013 as a subjective dataset with official multi-subject MOS and sample variance per image, then deliberately corrupts those labels to simulate a single-subject regime [2404.19666]. The biased-label process selects exactly one subject’s score per image when per-subject data are available, or otherwise draws a single sample from \(\mathcal{N}(\mathrm{MOS}, 0.2^2)\). The resulting noisy score \(y_n\) becomes the initialization \(u^0(x_n)\), and the original multi-annotator MOS is retained as held-out ground truth for evaluation. The contrast-assessment work reports a Dynamic-Reference Absolute Category Rating (DR-ACR) protocol, in which each image is shown with a randomly selected high-quality anchor image for the same scene, and up to 188 observers assign a five-point rating from “bad” to “excellent”; the ratings are then averaged into MOS [2509.21967]. Before training, MOS is Z-score normalized so that targets have zero mean and unit variance, and predictions are mapped back to the original MOS scale for evaluation. Taken together, these reports show that CID2013 functions not merely as an image collection but as a family of supervision regimes. The target variable may be a normalized MOS in \([0,100]\), a laboratory MOS with variance information, or a Z-scored MOS derived from DR-ACR. ## 3. Blind image-sharpness assessment with KANs In the KAN study, CID2013 is used for blind image-sharpness assessment with two feature regimes: 15 mid-level hand-crafted indicators and 2048 high-level deep features from a pre-trained ResNet50 [2409.07762]. Each image \(I_i\) is mapped either to \(x_{ij}=\eta_j(I_i)\), where the \(\eta_j\) are 15 sharpness or blurriness indicators spanning gradient-based, entropy-based, frequency-domain, and edge-width measures, or to the final fully connected activations \(f(I_i)\) of ResNet50. The KAN formulation replaces scalar edge weights with learnable univariate mappings \(\phi_{l,j,i}\), with forward propagation defined as $$ x_{l+1,j} = \sum_{i=1}^{n_l} \phi_{l,j,i}(x_{l,i}), $$ and $$ \mathrm{KAN}(x) = (\Phi_{L-1}\circ\cdots\circ\Phi_0)(x). $$ The paper evaluates TaylorKAN, BSRBF-KAN, ChebyKAN, HermiteKAN, JacobiKAN, and WavKAN. TaylorKAN is the proposed variant, with edge functions modeled by truncated Taylor expansion around a learnable center \(a\): $$ \phi(x) \simeq \sum_{n=0}^{N} \frac{f^{(n)}(a)}{n!}(x-a)^n, $$ with \(N=2\) in the experiments. Training uses a random 70\%/15\%/15\% train-validation-test split for CID2013, MSE loss \(L=\|y-R(x)\|^2\), Adam with default PyTorch settings, up to 500 epochs, and early stopping with patience \(=20\) on validation MSE. For mid-level inputs the layer widths are \([15 \rightarrow 26 \rightarrow 18 \rightarrow 12 \rightarrow 1]\); for high-level inputs they are \([2048 \rightarrow 1536 \rightarrow 1024 \rightarrow 256 \rightarrow 128 \rightarrow 1]\). No further normalization is reported, and the high-level features are extracted from center-cropped \(224\times224\) ResNet50 input. Evaluation uses PLCC and SRCC, with PLCC computed after the five-parameter non-linear mapping $$ f(s)=q_1\left(\frac{1}{2}-\frac{1}{1+e^{q_2(s-q_3)}}\right)+q_4 s+q_5. $$ | Method | 15 mid-level features (PLCC/SRCC) | 2048 deep features (PLCC/SRCC) | |---|---|---| | SVR | 0.834 / 0.810 | 0.860 / 0.882 | | MLP | 0.808 / 0.791 | 0.796 / 0.825 | | BSRBF-KAN | 0.845 / 0.795 | 0.828 / 0.820 | | ChebyKAN | 0.808 / 0.826 | 0.630 / 0.665 | | HermiteKAN | 0.825 / 0.845 | 0.604 / 0.687 | | JacobiKAN | 0.808 / 0.775 | 0.596 / 0.605 | | WavKAN | 0.827 / 0.827 | 0.844 / 0.856 | | TaylorKAN | 0.871 / 0.851 | 0.788 / 0.780 | On CID2013, TaylorKAN is the best method when the 15 mid-level features are used, reaching PLCC \(=0.871\) and SRCC \(=0.851\). The paper further notes that the 15 mid-level BISA indicators are more suitable than generic deep features for this dataset, since all KAN variants yield higher PLCC and SRCC with mid-level inputs. For the high-dimensional ResNet50 features, the study attributes weaker KAN performance to error accumulation and overfitting unless regularization or post-processing such as feature rectification is added. It also states that the early-stopping regime and the \([15\rightarrow\cdots\rightarrow1]\) architecture are sufficient to avoid over-fitting on CID2013’s 474 samples, while finer \(\lambda\)-regularization, adaptive learning rates, or message-passing integration remain open directions. ## 4. Perceptual-similarity preprocessing of CID2013 subjective scores The PSP method reinterprets a subject’s score \(y\) on image \(x\) as a conditional probability over latent subconscious reference images \(x'\) [2404.19666]. The core model is $$ p(y|x) = \mathbb{E}_{x'|x}\,[\,p(y|x,x')\cdot p(x'|x)\,], $$ with $$ p(y|x,x') \sim \mathcal{N}(u(x') + S(x,x'), \sigma^2). $$ Here \(u(x')\) denotes the “true” quality of the reference and \(S(x,x')\) is a perceptual residual derived from similarity. The intractable reference distribution \(p(x'|x)\) is approximated by retaining only the \(K\) nearest neighbors under a perceptual feature metric, with equal weight. The paper fixes \(K=1\) for simplicity. Feature extraction reuses LPIPS with a ResNet-50 backbone to obtain a perceptual depth vector \(f(x)\) for each image. Similarity is measured by normalized dot product, $$ \mathrm{sim}(x_i,x_j)=\frac{\langle f(x_i), f(x_j)\rangle}{\|f(x_i)\|\,\|f(x_j)\|}, $$ and the top-\(K\) most similar images are stored in a neighbor dictionary. At iteration \(t\), the method selects the most similar reference \(x_{n'}\), estimates a residual \(S^t(x_n,x_{n'})\) with a small CNN regressor on the image pair, forms the raw update \(r_n^t=S^t(x_n,x_{n'})+u^t(x_{n'})\), and applies similarity-regularized exponential moving average: $$ u^{t+1}(x_n)=\alpha\,[S^{t}(x_n,x_n') + u^t(x_n')] + (1-\alpha)\,u^t(x_n). $$ A warm-up period \(T\) is used so that for \(t