Abstract:

Image quality assessment often takes human subjective evaluation as the final measurement standard. However, manual assessment is time-consuming and cumbersome, and it cannot be applied to real-time images or video sequences. In the quality assessment system, therefore, a predictive image quality algorithm that aims to imitate human subjectivity has important value. In response to the above problems, this paper designs a convolutional neural network for local image quality assessment, which forms a more effective image quality assessment model by integrating feature learning and regression into an optimization process. According to human visual habits, this paper uses the eye tracker’s viewpoint distribution map to generate a weight pool based on visual importance. Experiments show that the design of the weight pool can achieve the best overall performance. The original image is low-pass filtered and down-sampled. The weight coefficient decay strategy is used in the down-sampling process, and the multi-scale image is used for weighted quality comprehensive assessment. The results prove that the multi-scale assessment method effectively improves the assessment model. The method in this paper can achieve excellent performance on the LIVE database and has good generalization ability.

Key words: image quality assessment, weight pool, receptive field, multi-scale assessment

摘要：

图像质量评估往往以人类的主观评估为最终衡量标准，然而人工评估耗时繁琐，又无法应用在对图像或者视频序列进行实时质量评估的系统中，因此一种旨在模仿人类主观性的预测图像质量算法具有重要的价值。针对上述问题，设计了一种用于局部图像质量评估的卷积神经网络，通过将特征学习和回归都集成到一个优化过程中，从而形成一种更有效的图像质量评估模型。根据人类的视觉习惯，利用眼动仪的视点分布图生成基于视觉重要性的权重池，利用高斯比例混合模型构造基于图像信息内容的权重池，实验证明权重池的设计可以获得最佳的整体性能。对原始图像进行低通滤波和下采样，下采样过程中采用权重系数衰退策略，利用多尺度的图像进行加权质量综合评估，实验结果证明多尺度评估方式有效地改进了评估模型。提出的方法在LIVE等数据库上可以达到优秀的性能，且具有不错的泛化能力。

关键词: 图像质量评估, 权重池, 感受野, 多尺度评估