CCID2014: Contrast Change Image Database

• CCID2014 is a no-reference contrast-distortion dataset comprising 655 synthetically altered images derived from 15 natural references, used for subjective quality assessment.
• It employs five distinct contrast operators, including gamma adjustments and mean shifts, to simulate diverse perceptual image alterations.
• Both handcrafted methods (like MDM) and deep-learning approaches (such as a customized EfficientNet-B0) have shown high correlation with human MOS, underscoring its benchmark value.

CCID2014, the Contrast Changed Image Database 2014, is a purpose-built no-reference contrast-distortion dataset for blind image quality assessment. It was introduced as a contrast-specific benchmark and is used to evaluate both handcrafted and deep no-reference image quality assessment models on synthetically altered natural images. In the conventions used by recent no-reference studies, the dataset is typically treated as a corpus of 655 distorted images with associated human Mean Opinion Scores (MOS), although the full collection includes 15 original reference images in addition to those distortions (Nafchi et al., 2018, Joloudari et al., 26 Sep 2025).

1. Dataset composition and scope

CCID2014 was introduced in “cyber2015” as a dataset specifically designed for no-reference assessment of contrast distortions. The dataset contains 655 distorted images generated from natural reference images. A later benchmark description states that 15 high-quality natural images are each synthetically re-rendered to produce a total of 670 images, comprising 15 originals and 655 distorted examples (Joloudari et al., 26 Sep 2025).

A common point of confusion is whether CCID2014 contains 655 or 670 images. Both counts are correct under different conventions. The count of 655 refers to the distorted subset used in no-reference experiments, while the count of 670 includes the original reference images. In no-reference quality prediction pipelines, the 15 originals are discarded and only the distorted images are used as regression targets (Joloudari et al., 26 Sep 2025).

The distorted images are reported as 8-bit grayscale at approximately $512\times384$ resolution. This resolution is also material to runtime analyses, because methods evaluated on CCID2014 often exploit down-sampling rules keyed to the image dimensions (Nafchi et al., 2018).

2. Distortion operators and generation process

CCID2014 is organized around five distinct contrast-distortion operators, each applied with multiple parameter settings. The reported operators are:

• Gamma-transfer (power-law) variations: exemplified by exponents $q \in \{0.2,0.4,\ldots,2.5\}$.
• Convex and concave arc-shaped mappings.
• Cubic and logistic-curve mappings.
• Mean-shifting: a global additive offset $\Delta$ within $[-50 \ldots +50]$.
• Compound mapping: a composite transformation that mixes power-law and offset.

These operators define CCID2014 as a contrast-distortion benchmark rather than a generic natural-scene distortion database. The dataset is therefore particularly suited to methods that model luminance redistribution, global tone remapping, and related perceptual effects. A later deep-learning description characterizes the synthetic alterations as “principally gamma corrections of increasing severity,” which is consistent with gamma variation being a central component, but the fuller operator list shows that the dataset spans a broader family of contrast-altering mappings (Joloudari et al., 26 Sep 2025).

The inclusion of both monotonic remappings and additive mean shifts is significant because the distortions do not all act through the same perceptual mechanism. This suggests that CCID2014 can probe whether a no-reference metric is sensitive only to histogram spread or whether it can also capture more structured tonal transformations.

3. Subjective annotation and target scores

For each distorted image, CCID2014 provides a Mean Opinion Score in $[0,100]$. The MOS is obtained by averaging the responses of multiple human observers. One description specifies that the subjective study followed ITU-R BT.500 recommendations, with observers viewing randomized image sequences and rating perceived quality on a continuous scale; another describes the protocol as Absolute Category Rating (Nafchi et al., 2018, Joloudari et al., 26 Sep 2025).

The MOS labels make CCID2014 a supervised regression benchmark for no-reference image quality assessment. In the 2025 deep-learning benchmark, the 15 original reference images are removed and the remaining 655 distorted images are assigned normalized MOS targets using a reported $Z$-score transformation before training (Joloudari et al., 26 Sep 2025).

Because the annotations are human-perceptual rather than analytically derived, CCID2014 is used to assess alignment between model predictions and subjective quality judgments. The standard performance measures reported on this dataset are Pearson Linear Correlation Coefficient (PLCC) and Spearman Rank Order Correlation Coefficient (SRCC), reflecting linear agreement after monotonic mapping and rank consistency, respectively (Joloudari et al., 26 Sep 2025).

4. Handcrafted no-reference modeling: the MDM pipeline

A prominent handcrafted baseline evaluated on CCID2014 is the Minkowski Distance based Metric (MDM), introduced for no-reference quality assessment of contrast distorted images. In this formulation, a distorted image $D$ is treated as a vector of $N$ pixels, $D=\{D_i\}_{i=1}^N$, and the mean of the power-law transformed pixels is defined as $\mu_q=\mathrm{mean}(D^q)$. The $q \in \{0.2,0.4,\ldots,2.5\}$0-order Minkowski deviation is then

$q \in \{0.2,0.4,\ldots,2.5\}$1

For dynamic-range stability, the proposed feature applies a fourth root:

$q \in \{0.2,0.4,\ldots,2.5\}$2

In the reported experiments, $q \in \{0.2,0.4,\ldots,2.5\}$3 and $q \in \{0.2,0.4,\ldots,2.5\}$4 are used. The same construction is also applied to the pixel complement $q \in \{0.2,0.4,\ldots,2.5\}$5, yielding a second feature. The third feature is the image entropy,

$q \in \{0.2,0.4,\ldots,2.5\}$6

where $q \in \{0.2,0.4,\ldots,2.5\}$7 is the empirical probability of gray level $q \in \{0.2,0.4,\ldots,2.5\}$8. The resulting feature vector is

$q \in \{0.2,0.4,\ldots,2.5\}$9

On CCID2014, these three features are fed to a Support-Vector Regressor with RBF kernel to predict MOS. The reported protocol uses an 80/20 content-separated split, with 80% of contents randomly chosen for training and 20% for testing, and no content overlap. After five-parameter logistic fitting, the held-out test performance is PLCC $\Delta$0 and SROCC $\Delta$1 (Nafchi et al., 2018).

The same three-feature representation can also be used for distortion-type classification via a Support-Vector Classifier. However, the exact multi-class CCID2014 classification accuracy was not tabulated in the cited work. The published description notes that analogous two-class tests on TID2013 exceed 94% accuracy with an 80/20 split and states an expectation of CCID2014 multi-class accuracy in excess of 80%. This is best treated as an informed expectation rather than a reported CCID2014 benchmark result.

5. Deep-learning benchmark: customized EfficientNet-B0

A later no-reference contrast assessment study uses CCID2014 as one of two benchmark datasets, alongside CID2013, and reports state-of-the-art no-reference performance on CCID2014 with a customized EfficientNet-B0 model (Joloudari et al., 26 Sep 2025). The model uses an ImageNet-pretrained EfficientNet-B0 backbone with compound scaling

$\Delta$2

EfficientNet-B0’s MBConv blocks extract a 1280-dimensional feature vector from each $\Delta$3 input. A contrast-aware regression head is attached on top, consisting of three fully connected layers with ReLU activations and dropout $\Delta$4 on the hidden layers:

$\Delta$5

with layer dimensions $\Delta$6. Training minimizes Mean Squared Error between predicted and ground-truth MOS.

The training protocol includes online augmentation targeted at tonal robustness: random horizontal flip with 50% chance, random rotation of $\Delta$7, color jitter with brightness $\Delta$8, contrast $\Delta$9, saturation $[-50 \ldots +50]$0, hue $[-50 \ldots +50]$1, and random zoom and shifts up to $[-50 \ldots +50]$2. Inputs are resized to $[-50 \ldots +50]$3, converted to tensors, and normalized with ImageNet $[-50 \ldots +50]$4 and $[-50 \ldots +50]$5. Optimization uses AdamW with initial learning rate $[-50 \ldots +50]$6, weight decay $[-50 \ldots +50]$7, batch size 32, 50 epochs, and a ReduceLROnPlateau scheduler. The backbone is initially frozen and then unfrozen if validation plateaus.

On CCID2014, the customized EfficientNet-B0 achieves PLCC $[-50 \ldots +50]$8 and SRCC $[-50 \ldots +50]$9, outperforming the compared no-reference baselines reported in the same study (Joloudari et al., 26 Sep 2025). The accompanying interpretation attributes the result to compound scaling, the contrast-aware regression head, targeted augmentation, transfer learning with fine-tuning, and optimization against perceptual MOS.

6. Computational characteristics and benchmark significance

The MDM study emphasizes computational efficiency on CCID2014. The implementation is reported in MATLAB 2013b on a Core-i7 3.4 GHz machine with 16 GB RAM. Images may be down-sampled by

$[0,100]$0

For CCID2014 images at approximately $[0,100]$1, this yields $[0,100]$2, halving each dimension. On a $[0,100]$3 input, MDM requires 5.23 ms per image; PSNR, a reference-based metric, requires 5.61 ms; and the next fastest no-reference competitor, NSS, requires approximately 23.9 ms. The same report states that all 655 CCID2014 distorted images are processed in under 4 s, and that the power-law and Minkowski exponentiations are implemented in $[0,100]$4 and $[0,100]$5 time via exponentiation-by-squaring (Nafchi et al., 2018).

The later EfficientNet-B0 study frames CCID2014 as a benchmark on which lightweight learned models can also deliver strong performance suitable for real-time and resource-constrained applications (Joloudari et al., 26 Sep 2025). Taken together, these results position CCID2014 as a dataset that supports two distinct lines of work: low-complexity feature-engineered no-reference metrics and compact deep architectures adapted for perceptual contrast assessment.

The comparative figures reported for CCID2014 illustrate this benchmarking role:

Method	PLCC	SRCC
RIQMC	0.8726	0.8465
RCIQM	0.8845	0.8565
NIQMC	0.9250	0.9040
MDM	0.8719	0.8368
Customized EfficientNet-B0	0.9286	0.9178

Within the reported comparisons, RIQMC and RCIQM are reduced-reference methods, while NIQMC, MDM, and the customized EfficientNet-B0 are no-reference methods (Joloudari et al., 26 Sep 2025). This suggests that CCID2014 is useful not only for absolute performance ranking but also for studying how different information regimes—reduced-reference versus no-reference, handcrafted versus learned—behave under controlled contrast distortions.

As a research object, CCID2014 is narrower than general-purpose IQA databases because it targets contrast alteration specifically. That specialization is precisely what makes it valuable in studies of blind contrast quality assessment: it isolates a well-defined distortion family, provides human MOS labels, and has been used to test both compact statistical descriptors and end-to-end convolutional regressors under reproducible protocols.