---
datasets:
- xiaoqi-wang/miqd-2.5m
---
Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models
[](https://huggingface.co/datasets/xiaoqi-wang/miqd-2.5m)
[](https://arxiv.org/abs/2508.19850)
[](https://huggingface.co/xiaoqi-wang/miqa)
[](https://github.com/XiaoqiWang/MIQA)
[δΈζ](README_CN.md) | [English](README.md) | [Model](https://huggingface.co/xiaoqi-wang/miqa) | [Dataset](https://huggingface.co/datasets/xiaoqi-wang/miqd-2.5m)
π― Project Overview
- π€ Machine-Centric: We bypass human perception to evaluate images from the perspective of the deep learning models that use them.
- π Task-Driven Metrics: Directly measure how degradations like blur, noise, or compression artifacts impact the performance of downstream vision tasks.
- π‘ A New Paradigm: MIQA offers a new lens for optimizing image processing pipelines where machines make the final decision.
## β¨ Does MIQA Work?
Performance improvement across tasks when filtering low-quality images using MIQA scores
ποΈ Key Results
Our results provide clear evidence of MIQA's effectiveness across three representative computer vision tasks: classification, detection, and segmentation.
The framework consistently identifies images that degrade model performance. By filtering these detrimental samples, MIQA directly leads to improved outcomes and demonstrates the universal utility of a machine-centric approach. This transforms quality assessment from a passive metric into a proactive tool, safeguarding downstream models against the unpredictable image quality of real-world conditions and ensuring robust performance when it matters most.
## π οΈ Installation Guide
#### Step 1: Install Dependencies
To get started, you'll need to install two essential libraries: **mmcv** and **mmsegmentation**.
Install mmcv and mmsegmentation
* For the latest version of **mmsegmentation**, follow the installation guide here:
[MMsegmentation Installation Guide](https://mmsegmentation.readthedocs.io/en/main/get_started.html)
* Alternatively, you can install a specific version of **mmsegmentation** based on your CUDA and PyTorch versions. You can find the version compatibility details here:
[MMCV Installation Guide](https://mmcv.readthedocs.io/zh-cn/latest/get_started/installation.html)
#### Step 2: Handle CUDA Version Compatibility
If your CUDA version is relatively high, such as 12.7 or higher, you might encounter a version mismatch with **mmcv**. In this case, you may need to install a compatible version of **mmcv**.
Install a compatible version of mmcv
For example, if you need a specific version of **mmcv**, you can uninstall the existing versions and install a compatible one as follows:
```bash
pip uninstall mmcv mmcv-full -y
mim install "mmcv>=2.0.0rc4,<2.2.0" # The version specified here is just an example. You should choose a version that is compatible with your CUDA and PyTorch setup.*
```
#### Step 3: Install Required Libraries
```bash
pip install -r requirements.txt
```
## π¦ Model Weights & Performance
### Where things live
### Naming & cache
Checkpoint pattern on the Hub: `miqa_ra_miqa_{cls|det|ins}_{composite|consistency|accuracy}_metric.pth.tar`
Examples:
- `miqa_ra_miqa_cls_composite_metric.pth.tar`
- `miqa_ra_miqa_det_consistency_metric.pth.tar`
- `miqa_ra_miqa_ins_accuracy_metric.pth.tar`
On first run, `huggingface_hub` downloads into `models/checkpoints/{composite|consistency|accuracy}_metric/`.
## π Quick Start
### Assess a Single Image
Run MIQA inference for a single image using command-line interface:
```
# Evaluate a single image for classification-oriented MIQA
python img_inference.py --input path/to/image.jpg --task cls --model ra_miqa
```
### Evaluate a Directory of Images
Process all images within a directory
```
# Assess all images in a directory (e.g., detection-oriented MIQA)
python img_inference.py --input ./assets/demo_images/coco_demo --task det --model ra_miqa
```
### Save Results and Visualizations
To save outputs and generate visualized results:
```
# Save the predicted scores and visualization for a single image
python img_inference.py --input path/to/image.jpg --task cls --model ra_miqa --save-results --visualize
# Save batch results and generate visualization for a directory
python img_inference.py --input ./assets/demo_images/imagenet_demo --task ins --save-results --visualize
```
### π¬ Video Assessment
Video Quality Assessment offers two workflows: **(1) Frame-by-Frame Annotation**: Generates fully annotated videos for detailed visual inspection. Suitable for demos and qualitative analysis but computationally intensive.
**(2) Selective Sampling & Aggregation**: Samples frames to produce plots and structured data (.json) for efficient, quantitative analysis. Ideal for batch processing and reporting.
#### Analyze a Single Video (**Frame-by-Frame Annotation**)
Run MIQA video inference for one video and save the annotated output.
```bash
# Evaluate a single video using RA-MIQA (classification-oriented MIQA)
python video_annotator_inference.py --input assets/demo_video/brightness_distorted.mp4 --task cls --model ra_miqa
```
#### Evaluate a Directory of Videos (**Frame-by-Frame Annotation**)
Process all videos within a given folder:
```bash
# Assess all videos in a directory for object detection-oriented MIQA
python video_annotator_inference.py --input assets/demo_video/ --task det --model ra_miqa
```
The primary output is a new `.mp4` video file. This video shows the original footage playing alongside a dynamic side panel that displays the real-time quality score and a line chart that grows as the video progresses.
π₯ Example: Frame-wise MIQA Predictions on Videos
| Brightness Variation |
Compression Artifacts |
Minimal Perceptual Distortion |
|
|
|
#### Analyze a Single Video (**Selective Sampling & Aggregation**)
For efficient, quantitative analysis, this script samples frames from the video instead of processing all of them. It is significantly faster and designed for generating analytical reports.
```bash
# Analyze a video, sample frames, and create a dual-granularity plot
python video_analytics_inference.py --input assets/demo_video/gaussian_distorted.mp4 --task ins --visualize --viz-granularity both
```
#### Evaluate a Directory of Videos (**Selective Sampling & Aggregation**)
This workflow is highly optimized for batch processing.
```bash
# Analyze all videos in a directory, sampling 120 frames from each
python video_analytics_inference.py --input assets/demo_video/ --task det --video-frames 120 --visualize
python video_analytics_inference.py --input assets/demo_video/jpeg_distorted.mp4 --task det --visualize --viz-granularity both
# viz-granularity both : Specifies the type of plot to generate. 'composite' creates a comprehensive, side-by-side comparison chart showing:
#1. The raw, frame-level quality scores. 2. The smoothed, per-second average quality scores.
```
This process **does not create a new video**. Instead, it generates two key outputs for each video analyzed:
1. A **`.png` image**: A detailed time-series plot showing the quality score fluctuation over the video's duration.
2. A **`.json` file**: A structured data file containing per-second aggregated scores, overall statistics (average, min, max, std. dev), and video metadata.
π₯ Example: Frame-wise MIQA Predictions on Videos
| Brightness Variation |
Compression Artifacts |
Minimal Perceptual Distortion |
 |
 |
 |
## π Training and Evaluation
### Training
```bash
CUDA_VISIBLE_DEVICES=0,1 python train.py \
--dataset 'miqa_cls' \
--path_miqa_cls 'path/to/datasets_miqa_cls' \
--train_split_file '../data/dataset_splitting/miqa_cls_train.csv' \
--val_split_file '../data/dataset_splitting//miqa_cls_val.csv' \
--metric_type 'composite' --loss_name 'mse' --is_two_transform \
-a 'RA-MIQA' --pretrained --transform_type 'simple_transform' \
-b 256 --epochs 5 --warmup_epochs 1 --validate_num 2 --lr 1e-4 \
--image_size 288 --crop_size 224 --workers 8 -p 100 \
--multiprocessing-distributed --world-size 1 --rank 0
```
More training scripts are available in the "**scripts**" directory.
### Evaluation on Standard Benchmarks
```bash
# Evaluate on miqa_cls val set
python evaluate.py --model_name ra_miqa --train_dataset cls --test_dataset cls --metric_type composite
# Cross-dataset evaluation: evaluate the RA-MIQA model trained on miqa_cls dataset and tested on miqa_det dataset
python evaluate.py --model_name ra_miqa --train_dataset cls --test_dataset det --metric_type composite
```
## π Benchmarks
Tabel 1: Performance Benchmark on Composite Score
| Category |
Method |
Image Classification |
Object Detection |
Instance Segmentation |
| SRCC β |
PLCC β |
KRCC β |
RMSE β |
SRCC β |
PLCC β |
KRCC β |
RMSE β |
SRCC β |
PLCC β |
KRCC β |
RMSE β |
| HVS-based |
PSNR |
0.2388 |
0.2292 |
0.1661 |
0.2928 |
0.3176 |
0.3456 |
0.2148 |
0.2660 |
0.3242 |
0.3530 |
0.2196 |
0.2553 |
| SSIM |
0.3027 |
0.2956 |
0.2119 |
0.2874 |
0.4390 |
0.4505 |
0.3011 |
0.2531 |
0.4391 |
0.4512 |
0.3011 |
0.2435 |
| VSI |
0.3592 |
0.3520 |
0.2520 |
0.2816 |
0.4874 |
0.4940 |
0.3355 |
0.2465 |
0.4919 |
0.4985 |
0.3392 |
0.2365 |
| LPIPS |
0.3214 |
0.3280 |
0.2258 |
0.2842 |
0.5264 |
0.5376 |
0.3697 |
0.2390 |
0.5342 |
0.5453 |
0.3754 |
0.2287 |
| DISTS |
0.3878 |
0.3804 |
0.2724 |
0.2782 |
0.5266 |
0.5352 |
0.3659 |
0.2395 |
0.5363 |
0.5450 |
0.3738 |
0.2288 |
| HyperIQA |
0.2496 |
0.2279 |
0.1741 |
0.2929 |
0.4462 |
0.4463 |
0.3031 |
0.2537 |
0.4456 |
0.4518 |
0.3031 |
0.2434 |
| MANIQA |
0.3403 |
0.3255 |
0.2387 |
0.2844 |
0.4574 |
0.4617 |
0.3124 |
0.2515 |
0.4636 |
0.4680 |
0.3176 |
0.2411 |
|
| Machine-based |
ResNet-18 |
0.5131 |
0.5427 |
0.3715 |
0.2527 |
0.7541 |
0.7734 |
0.5625 |
0.1797 |
0.7582 |
0.7790 |
0.5674 |
0.1711 |
| ResNet-50 |
0.5581 |
0.5797 |
0.4062 |
0.2451 |
0.7743 |
0.7925 |
0.5824 |
0.1729 |
0.7729 |
0.7933 |
0.5826 |
0.1661 |
| EfficientNet-b1 |
0.5901 |
0.6130 |
0.4320 |
0.2377 |
0.7766 |
0.7950 |
0.5859 |
0.1720 |
0.7808 |
0.7999 |
0.5918 |
0.1637 |
| EfficientNet-b5 |
0.6330 |
0.6440 |
0.4680 |
0.2301 |
0.7866 |
0.8041 |
0.5971 |
0.1685 |
0.7899 |
0.8074 |
0.6013 |
0.1610 |
| ViT-small |
0.5998 |
0.6161 |
0.4407 |
0.2370 |
0.7992 |
0.8142 |
0.6099 |
0.1646 |
0.7968 |
0.8139 |
0.6083 |
0.1585 |
| RA-MIQA |
0.7003
|
0.6989
|
0.5255
|
0.2152
|
0.8125
|
0.8264
|
0.6263
|
0.1596
|
0.8188
|
0.8340
|
0.6333
|
0.1505
|
Table 2: Consistency & Accuracy Score Benchmark
| Method |
Image Classification |
Object Detection |
Instance Segmentation |
| Accuracy Score |
Consistency Score |
Accuracy Score |
Consistency Score |
Accuracy Score |
Consistency Score |
| SRCC β |
PLCC β |
RMSE β |
SRCC β |
PLCC β |
RMSE β |
SRCC β |
PLCC β |
RMSE β |
SRCC β |
PLCC β |
RMSE β |
SRCC β |
PLCC β |
RMSE β |
SRCC β |
PLCC β |
RMSE β |
| HVS-based Methods |
| PSNR |
0.2034 |
0.1620 |
0.3541 |
0.2927 |
0.2812 |
0.2692 |
0.2234 |
0.2449 |
0.2747 |
0.3712 |
0.3933 |
0.2839 |
0.2182 |
0.2398 |
0.2616 |
0.3796 |
0.4061 |
0.2770 |
| SSIM |
0.2529 |
0.2101 |
0.3509 |
0.3740 |
0.3663 |
0.2610 |
0.3434 |
0.3419 |
0.2662 |
0.5128 |
0.5130 |
0.2651 |
0.3271 |
0.3284 |
0.2545 |
0.5174 |
0.5204 |
0.2589 |
| VSI |
0.3020 |
0.2515 |
0.3473 |
0.4392 |
0.4336 |
0.2528 |
0.3799 |
0.3685 |
0.2634 |
0.5700 |
0.5571 |
0.2565 |
0.3703 |
0.3645 |
0.2509 |
0.5757 |
0.5749 |
0.2481 |
| LPIPS |
0.2680 |
0.2355 |
0.3488 |
0.3927 |
0.4032 |
0.2567 |
0.4064 |
0.3987 |
0.2598 |
0.6196 |
0.6232 |
0.2415 |
0.3972 |
0.3941 |
0.2476 |
0.6300 |
0.6344 |
0.2344 |
| DISTS |
0.3291 |
0.2768 |
0.3448 |
0.4683 |
0.4628 |
0.2487 |
0.4089 |
0.3999 |
0.2597 |
0.6174 |
0.6178 |
0.2429 |
0.4069 |
0.4012 |
0.2468 |
0.6255 |
0.6270 |
0.2362 |
| HyperIQA |
0.2100 |
0.1649 |
0.3540 |
0.2966 |
0.2777 |
0.2695 |
0.3646 |
0.3545 |
0.2649 |
0.5009 |
0.4943 |
0.2684 |
0.3486 |
0.3442 |
0.2530 |
0.5056 |
0.4995 |
0.2626 |
| MANIQA |
0.2924 |
0.2435 |
0.3481 |
0.3963 |
0.3870 |
0.2587 |
0.3839 |
0.3823 |
0.2618 |
0.4991 |
0.4975 |
0.2679 |
0.3755 |
0.3749 |
0.2498 |
0.5096 |
0.5098 |
0.2608 |
| Machine-based Methods |
| ResNet-50 |
0.4734 |
0.4411 |
0.3221 |
0.5989 |
0.6551 |
0.2119 |
0.6955 |
0.6898 |
0.2051 |
0.8252 |
0.8457 |
0.1648 |
0.6863 |
0.6847 |
0.1964 |
0.8320 |
0.8480 |
0.1607 |
| EfficientNet-b5 |
0.5586 |
0.5149 |
0.3076 |
0.6774 |
0.7168 |
0.1956 |
0.7042 |
0.6991 |
0.2026 |
0.8353 |
0.8530 |
0.1612 |
0.6933 |
0.6949 |
0.1938 |
0.8419 |
0.8564 |
0.1565 |
| ViT-small |
0.5788 |
0.5197 |
0.3066 |
0.6798 |
0.7189 |
0.1950 |
0.7121 |
0.7052 |
0.2008 |
0.8459 |
0.8620 |
0.1566 |
0.7168 |
0.7146 |
0.1885 |
0.8487 |
0.8616 |
0.1539 |
| RA-MIQA |
0.6573
|
0.5823
|
0.2917
|
0.7707
|
0.7866
|
0.1732
|
0.7448
|
0.7370
|
0.1915
|
0.8526
|
0.8692
|
0.1527
|
0.7363
|
0.7327
|
0.1834
|
0.8632
|
0.8756
|
0.1464
|
## π Citation
If you find this work useful in your research, please consider citing:
```bibtex
@article{wang2025miqa,
title={Image Quality Assessment for Machines: Paradigm, Large-scale Database, and Models},
author={Wang, Xiaoqi and Zhang, Yun and Lin, Weisi},
journal={arXiv preprint arXiv:2508.19850},
year={2025}
}
```
## π§ Contact
- **Project Maintainer**: [Xiaoqi Wang](mailto:wangxq79@mail2.sysu.edu.cn)
- **Issues**: Please use [GitHub Issues](https://github.com/XiaoqiWang/MIQA/issues) for bug reports and feature requests 
