کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

Proxy-Supervised Learning for No-Reference Screen Content Quality Assessment via FISH-based Aggregation

نویسندگان :

Zohreh Yarahmady¹ َAzadeh Mansouri²

1- خوارزمی 2- خوارزمی

کلمات کلیدی :

No-Reference Image Quality Assessment،FISH Algorithm،Screen Content Images،Generalization،Proxy Labels،Cross-Dataset Evaluation،Convolutional Neural Network،Patch-Based Processing

چکیده :

Image Quality Assessment (IQA) for Screen Content Images (SCIs) remains a challenging task due to the distinct statistical characteristics of SCIs compared to natural images, which limits the effectiveness of conventional IQA methods. Additionally, the scarcity of human-annotated quality labels and the impracticality of assuming constant access to reference images in real-world scenarios motivate the development of robust No-Reference (NR) IQA techniques. In this paper, we propose a novel NR-SCIQA method that takes advantage of patch-based processing and proxy labels to overcome data limitations and improve generalization. Each image is divided into patches, and full-reference quality scores generated by the wavelet-based WS-HV algorithm serve as proxy labels for training a convolutional neural network (CNN). The model predicts patch-level quality, which is then aggregated into a global score using the FISH sharpness-based weighting strategy. Experimental results on the SIQAD and SCID datasets demonstrate that our method achieves state-of-the-art performance, particularly in cross-dataset evaluation and when handling both seen and unseen distortions. The proposed approach exhibits strong generalization capabilities, outperforming existing methods in terms of SROCC, PLCC, and RMSE metrics.