Opto-Electronic Engineering, Volume. 52, Issue 4, 240309(2025)
Multi-task attention mechanism based no reference quality assessment algorithm for screen content images
Figures & Tables(18)
Fig. 3. Structure of group-wise attention mechanism with spatial shifts
Fig. 4. Structure of asymmetric convolutional channel attention mechanism
Fig. 6. Variation curves of PLCC, SRCC, and loss values obtained on the SCID dataset. (a) Variation curves of PLCC and SRCC; (b) Variation curve of the loss value
Fig. 7. Variation curves of PLCC, SRCC, and loss values obtained on the SIQAD dataset. (a) Variation curves of PLCC and SRCC; (b) Variation curve of the loss value
Fig. 8. Distorted screen content image. (a) Reference image; (b) SCI33_5_3.bmp; (c) SCI33_5_4.bmp; (d) SCI33_5_5.bmp
Table 1. Typical methods of screen content image quality assessment
View tableView in ArticleTable 1. Typical methods of screen content image quality assessment
Category Method Type Feature The first category SPQA[ 6 ]FR Brightness and sharpness ESIM[ 9 ]FR Edge contrast MSDL[ 19 ]FR Feature extraction using log gabor filters BLIQUP-SCI[ 7 ]NR Natural scene statistics features and local texture Yang et al.[ 8 ]NR The amplitude, variance, entropy, and edge structure of wavelet coefficients Huang et al.[ 20 ]RR Oriented histogram, local discrete cosine transform coefficients, and gradient of amplitude in color channels The second category SR-CNN[ 10 ]FR Multi-level CNN features QODCNN[ 12 ]FR/NR CNN features Gao et al.[ 15 ]NR CNN features Zhang et al.[ 16 ]NR CNN features MIC-CNN[ 13 ]NR CNN features SIQA-DF-II[ 11 ]NR CNN features RIQA[ 14 ]NR CNN features DAMC[ 21 ]FR CNN features MTDL[ 17 ]NR CNN features
Table 2. Group-based spatial shift operation
View tableView in ArticleTable 2. Group-based spatial shift operation
X nSpatial shift n=1 Shift1：move tensor x1 down by one pixel vertically and right by one pixel horizontally n=2 Shift2：move tensor x2 down by two pixels vertically and right by two pixels horizontally n=3 Shift3：move tensor x3 right by one pixel horizontally and down by one pixel vertically n=4 Without any processing
Table 3. Convolutional kernel size and padding method
View tableView in ArticleTable 3. Convolutional kernel size and padding method
Convolution layer Kernel size Padding size Conv0 5×5 2×2 Conv0_1 1×5 0×2 Conv0_2 5×1 2×0 Conv1_1 1×13 0×6 Conv1_2 13×1 6×0 Conv2_1 1×19 9×0 Conv2_2 19×1 0×9 Conv3 1×1 1×1
Table 4. Commonly used screen content image datasets
View tableView in ArticleTable 4. Commonly used screen content image datasets
Dataset Number of reference Number of distortion Distortion types count Distortion levels count Subjective score type SCID 40 1800 9 5 MOS SIQAD 20 980 7 7 DMOS
Table 5. Environmental configuration and parameters of the experiment
View tableView in ArticleTable 5. Environmental configuration and parameters of the experiment
Parameter Value Param count 307.94577 M Compilation Environment Python 3.7.0, Pytorch-GPU 1.13.1, and CUDA 11.3 CPU model Intel Core i7-13700 GPU model NVIDIA RTX 4090 Average time/epoch 90 s
Table 6. Performance comparison of various screen content image quality assessment algorithms
View tableView in ArticleTable 6. Performance comparison of various screen content image quality assessment algorithms
Type Method SCID SIQAD SRCC PLCC SRCC PLCC FR MIC-CNN[ 13 ]- - 0.9636 0.9669 ESIM[ 9 ]0.8478 0.8630 0.8632 0.8788 DAMC[ 21 ]0.9617 0.9617 0.9304 0.9373 SR-CNN[ 10 ]0.9400 0.9390 0.8943 0.9042 NR Yang et al.[ 8 ]0.7562 0.7867 0.8543 0.8738 QODCNN[ 12 ]0.8760 0.8820 0.8890 0.9010 RIQA[ 14 ]- 0.9000 0.9110 Zhang et al.[ 16 ]0.9050 0.9133 0.9242 0.9260 BLIQUP-SCI[ 7 ]- - 0.7990 0.7705 Yang et al.[ 8 ]0.7562 0.7867 0.8543 0.8738 SIQA-DF-II[ 11 ]- - 0.8880 0.9000 Gao et al.[ 15 ]0.8569 0.8613 0.8962 0.9000 MTDL[ 17 ]- - 0.9233 0.9248 DFSS-IQA[ 29 ]0.8146 0.8138 0.8820 0.8818 Zhang[ 30 ]0.9445 0.9433 0.8640 0.8889 MTA-SCI 0.9602 0.9609 0.9233 0.9294
Table 7. Predicted scores of the distorted screen content images
View tableView in ArticleTable 7. Predicted scores of the distorted screen content images
Image ID Prediction value MOS Normalized MOS SCI33_5_3.bmp 0.1067 36.7569 0.2711 SCI33_5_4.bmp 0.0653 25.4741 0.0619 SCI33_5_5.bmp 0.0658 25.6376 0.0650
Table 8. Impact of different numbers of groups on the MTA-SCI performance
View tableView in ArticleTable 8. Impact of different numbers of groups on the MTA-SCI performance
Group count PLCC SRCC k=2 0.9471 0.9520 k=3 0.9591 0.9587 k=4 0.9602 0.9609 k=5 0.9572 0.9487
Table 9. Impact of different asymmetric convolution kernel combinations on the performance of the MTA-SCI
View tableView in ArticleTable 9. Impact of different asymmetric convolution kernel combinations on the performance of the MTA-SCI
Kernel combination PLCC SRCC Kernal=3, 15, 19 0.9568 0.9585 Kernal=5, 7, 9 0.9569 0.9584 Kernal= 5, 13, 19 0.9602 0.9609 Kernal=7, 11, 21 0.9575 0.9563 Kernal=9, 17, 25 0.9581 0.9589 Kernal=15, 23, 27 0.8967 0.9005
Table 10. Impact of ILAM, DFM, and residual connection on the algorithm performance
View tableView in ArticleTable 10. Impact of ILAM, DFM, and residual connection on the algorithm performance
No. ILAM DFM RC PLCC SRCC 1 × × × 0.8792 0.8651 2 × √ × 0.8832 0.8796 3 √ × × 0.9481 0.9508 4 √ √ × 0.9571 0.9592 5 √ √ √ 0.9602 0.9609
Tools
Get Citation
Copy Citation Text
Ziyi Zhou, Wu Dong, Likun Lu, Qian Ma, Guopeng Hou, Erqing Zhang. Multi-task attention mechanism based no reference quality assessment algorithm for screen content images[J]. Opto-Electronic Engineering, 2025, 52(4): 240309
Paper Information
Category: Article
Received: Dec. 30, 2024
Accepted: Feb. 25, 2025
Published Online: Jun. 11, 2025
The Author Email: Wu Dong (董武)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20