基于多尺度卷积神经网络的无参考图像质量评价

Non-reference Image Quality Evaluation Based on Multi-scale Convolutional Neural Network

图像质量评价广泛应用于图像处理领域。提出了一种基于多尺度卷积神经网络(multi-scale convolutional neural network,MCN-Net)的无参考图像质量评价方法,该方法将卷积神经网络与迁移学习相结合。首先,通过使用Resnet-50网络与感知模块相结合的方式,有效提取满足人类感知的多尺度语义信息特征;然后利用自适应融合网络对局部语义信息进行特征融合;最后,将融合后的局部语义信息与全局语义信息进行连接,并输入到全连接回归网络实现图像质量预测。为了验证模型的有效性,分别在LIVE、KonIQ-10K,LIVEC数据集上做了性能对比试验,实验结果表明,所提模型的图像质量评价性能优于目前大多数主流方法,并且在真实失真数据集上的泛化性能更好,适合用于自然失真场景。

Image quality evaluation is widely applied in image processing field.In this paper,an evaluation method for non-reference image quality based on multi-scale convolutional neural network is proposed,which combines convolutional neural network with migration learning.Firstly,by using the combination of Resnet-50 network and perception module,the multi-scale semantic information features satisfying human perception system are extracted effectively Then,the features of the local semantic information are fused by the adaptive fusion network,and finally,the local semantic information after fusion is connected with the global semantic information,and input into the full-connection regression network to effectively evaluate the image quality.In order to verify the validity of the model,the performance comparison experiments are done on LIVE,Kaniq-10k and LIVEC data sets respectively.The experimental results show that the performance of the proposed model is better than most of the current mainstream algorithms,and the generalization performance on the true distortion data set is better,which is suitable for natural distortion scenes.