摘要： 图像质量评价模型通过提取并分析与人类视觉系统相一致的图像特征来达到评价图像质量的目的。随着深度学习技术的发展，很多基于深度学习的图像质量评价模型相继出现，但是多数模型在小数据量环境下容易出现过拟合问题。为此，通过对DIQaM_FR/NR模型进行改进，建立一种Res-DIQaM_FR/NR图像质量评价模型。采用迁移学习方法，利用预训练的ResNet50网络替代DIQaM_FR/NR的特征提取层以进行图像特征提取，使用全局平均池化层取代DIQaM_FR/NR中的全连接层FC-512从而对提取的特征进行回归学习。实验结果表明，该模型在降低DIQaM_FR/NR模型复杂度的同时能够深化其网络结构，在小数据量环境下也能较好地模拟人类视觉系统并准确评价图像质量。

关键词: 图像质量评价, 特征提取, 迁移学习, 深度学习, ResNet50网络

Abstract: Image quality assessment models evaluate image quality by extracting and analyzing image features which are consistent with the human visual system.In recent years,with the development of deep learning technologies,many image quality assessment models based on deep learning have emerged,but most of them are prone to over-fitting problem with a small scale of data.To address the problem,this paper establishes a Res-DIQaM_FR/NR image quality assessment model by improving the DIQaM_FR/NR model.The improved model uses the transfer learning method to replace the original feature extraction layers of DIQaM_FR/NR with the pre-trained ResNet50 network for image feature extraction.Also,the global average pooling layer is used to replace the fully connected layer FC-512 of DIQaM_FR/NR for regressive learning of extracted features.Experimental results show that the proposed model can reduce the complexity of DIQaM_FR/NR while deepening its network structure,and can successfully simulate the visual system of human beings with a small scale of data to accurately assess image quality.

Key words: image quality assessment, feature extraction, transfer learning, deep learning, ResNet50 network

中图分类号: 

• TP391