TY - GEN
T1 - SUR-Net
T2 - 11th International Conference on Quality of Multimedia Experience
AU - Fan, Chunling
AU - Lin, Hanhe
AU - Hosu, Vlad
AU - Zhang, Yun
AU - Jiang, Qingshan
AU - Hamzaoui, Raouf
AU - Saupe, Dietmar
N1 - Funding Information:
This work was supported in part by the NSFC under Grant 61871372, Guangdong NSF for Distinguished Young Scholar under Grant 2016A030306022, Guangdong Provincial Science and Technology Development under Grant 2017B010110014, Shenzhen International Collaborative Research Project under Grant GJHZ20170314155404913, Shenzhen Science and Technology Program under Grant JCYJ20170811160212033, Guangdong International Science and Technology Cooperative Research Project under Grant 2018A050506063, Membership of Youth Innovation Promotion Association, CAS under Grant 2018392, and Shenzhen Discipline Construction Project for Urban Computing and Data Intelligence. This work was also funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) Projektnummer 251654672 TRR 161 (Project A05).
Publisher Copyright:
© 2019 IEEE.
PY - 2019/6/24
Y1 - 2019/6/24
N2 - The Satisfied User Ratio (SUR) curve for a lossy image compression scheme, e.g., JPEG, characterizes the probability distribution of the Just Noticeable Difference (JND) level, the smallest distortion level that can be perceived by a subject. We propose the first deep learning approach to predict such SUR curves. Instead of the direct approach of regressing the SUR curve itself for a given reference image, our model is trained on pairs of images, original and compressed. Relying on a Siamese Convolutional Neural Network (CNN), feature pooling, a fully connected regression-head, and transfer learning, we achieved a good prediction performance. Experiments on the MCL-JCI dataset showed a mean Bhattacharyya distance between the predicted and the original JND distributions of only 0.072.
AB - The Satisfied User Ratio (SUR) curve for a lossy image compression scheme, e.g., JPEG, characterizes the probability distribution of the Just Noticeable Difference (JND) level, the smallest distortion level that can be perceived by a subject. We propose the first deep learning approach to predict such SUR curves. Instead of the direct approach of regressing the SUR curve itself for a given reference image, our model is trained on pairs of images, original and compressed. Relying on a Siamese Convolutional Neural Network (CNN), feature pooling, a fully connected regression-head, and transfer learning, we achieved a good prediction performance. Experiments on the MCL-JCI dataset showed a mean Bhattacharyya distance between the predicted and the original JND distributions of only 0.072.
KW - Convolutional Neural Network
KW - Deep Learning
KW - Just Noticeable Difference
KW - Satisfied User Ratio
UR - http://www.scopus.com/inward/record.url?scp=85068692063&partnerID=8YFLogxK
U2 - 10.1109/QoMEX.2019.8743204
DO - 10.1109/QoMEX.2019.8743204
M3 - Conference contribution
AN - SCOPUS:85068692063
SN - 978-1-5386-8213-5
T3 - 2019 11th International Conference on Quality of Multimedia Experience, QoMEX 2019
BT - 2019 11th International Conference on Quality of Multimedia Experience (QoMEX)
PB - IEEE
CY - Piscataway, NJ
Y2 - 5 June 2019 through 7 June 2019
ER -