Blind image quality prediction by exploiting multi-level deep representations

作者：

Highlights：

• We leverage the benefits introduced by very deep DNNs and the difficulty in training a very deep DNN model.

• We reason the image quality at each level of representation, to get the best of both the intermediate-level and high-level representations.

• The proposed method works remarkably well and is highly comparable to state-of-the-art BIQA methods, over various canonical datasets.

摘要

•We leverage the benefits introduced by very deep DNNs and the difficulty in training a very deep DNN model.•We reason the image quality at each level of representation, to get the best of both the intermediate-level and high-level representations.•The proposed method works remarkably well and is highly comparable to state-of-the-art BIQA methods, over various canonical datasets.

论文关键词：Image quality assessment,Deep learning,Convolutional Neural Networks (CNN),Multi-level deep representation,Support vector regression

论文评审过程：Received 2 May 2017, Revised 20 December 2017, Accepted 11 April 2018, Available online 13 April 2018, Version of Record 22 April 2018.

论文官网地址：https://doi.org/10.1016/j.patcog.2018.04.016