An integration of deep network with random forests framework for image quality assessment in real-time

Over the last decade, the generation and streaming of images by data providers have grown rapidly. Facial images, in particular, have gained notable interest because of their extensive applications in social media. Nevertheless, the large-scale sharing of images across these platforms presents considerable challenges and necessitates substantial computational resources for efficient processing. Additionally, images are prone to multiple types of distortions arising from processing, transmission, sharing, or other real-world conditions. Therefore, assessing image quality is crucial to ensure the reliable delivery of content, especially in cases where the original, undistorted version is unavailable. In this paper, we propose a novel, lightweight, and reference-free image quality assessment framework that integrates deep convolutional neural networks for feature extraction with an ensemble-based Random Forest model for quality prediction. Unlike existing approaches, the proposed method incorporates a compact, face-aware representation through a novel face alignment metric, enabling more accurate quality estimation in face-centric streaming scenarios while maintaining low computational complexity. For experimental validation, we employed the artificially distorted LIVE and TID2013 benchmark datasets. The results demonstrate that our method outperforms existing state-of-the-art techniques, achieving a Pearson Correlation Coefficient (PCC) of approximately 0.942 and a Spearman Rank Order Correlation Coefficient (SROCC) of approximately 0.931 when compared with human subjective scores. Moreover, the proposed method achieves a processing time reduction from 4.8 ms to 1.8 ms, enhancing overall computational efficiency.