1、VeNICE: A very deep neural network approach to no-reference image assessment.
1.1框架:总共包括5个group,group1:conv,conv,relu,maxpool;group2:conv,conv,relu,maxpool;group3:conv,conv,conv,relu,maxpool;group4:conv,conv,conv,relu,maxpool;group5:conv,conv,conv,relu,maxpool;卷积层的参数直接fine-tuneVGG16的参数。全连接层:49-512-4096-1。
1.2 参数设置:input:224*224;batch size:7;epochs:200;learning rate:0.0001
2、2016ICASSP:Blind image quality assessment formultiply distorted images via convolutional networks. 同济大学
2.1相比于2014年的CVPR(Le Kang),此篇论文增加了feature map的数量,增加了全连接的数量。
2.2 Conv + max pool,average pool,dense 2048, dense 2048, output.
2.3