With the emerging use of technology and screen-oriented applications in our daily life, screen content images have gained the same importance as natural scene images. This results in many natural-scene and screen-content blind image quality assessment (BIQA) models to evaluate the perceptual quality without any prior information regarding the reference image. Recently, patch-based techniques for image quality assessment (IQA) have shown promising results. As per our knowledge, no IQA technique in literature is available that can be equally effective for both natural-scene and screen-content images. In this work, we have proposed a deep architecture with a region proposal network (RPN) for blind natural-scene and screen-content image quality assessment, named DeepRPN-BIQA. The proposed architecture computes visual saliency using RPN to extract important regions having a high contribution towards the image quality. Important regions are extracted by utilizing the texture and edges of images by sliding the network over the extracted feature map from deep architectures i.e., VGGNet and ResNet. The regions proposed (RP) that overlap more than 60% are merged into one proposal and are called the region of interest (ROI). The overlap between RPs is computed using anchors having 3 different scales and aspect ratios. A local quality score is computed over each ROI and the total quality score is computed by taking the average of all the local quality scores. Experimental results show that the DeepRPN-BIQA shows a high correlation between mean observer score and predicted quality score and performs better than other models for screen content images, synthetically distorted images, images taken in real-life conditions using mobile phone cameras, and large scale image quality assessment database.

随着我们日常生活中不断涌现的技术和面向屏幕的应用程序的使用，屏幕内容图像已获得与自然场景图像相同的重要性。这导致许多自然场景和屏幕内容盲图像质量评估 (BIQA) 模型在没有任何关于参考图像的先验信息的情况下评估感知质量。最近，用于图像质量评估 (IQA) 的基于补丁的技术已显示出有希望的结果。据我们所知，文献中没有可用的 IQA 技术对自然场景和屏幕内容图像同样有效。在这项工作中，我们提出了一种带有区域提议网络 (RPN) 的深层架构，用于盲自然场景和屏幕内容图像质量评估，称为 DeepRPN-BIQA。所提出的架构使用 RPN 计算视觉显着性，以提取对图像质量有很大贡献的重要区域。通过在从深度架构（即 VGGNet 和 ResNet）提取的特征图上滑动网络，利用图像的纹理和边缘提取重要区域。重叠超过 60% 的提议区域 (RP) 合并为一个提议，称为感兴趣区域 (ROI)。RP 之间的重叠是使用具有 3 种不同比例和纵横比的锚点计算的。在每个 ROI 上计算局部质量分数，并通过取所有局部质量分数的平均值来计算总质量分数。