TY - JOUR
T1 - Image restoration from patch-based compressed sensing measurement
AU - Huang, Hua
AU - Nie, Guangtao
AU - Zheng, Yinqiang
AU - Fu, Ying
N1 - Publisher Copyright:
© 2019
PY - 2019/5/7
Y1 - 2019/5/7
N2 - A series of methods have been proposed to restore an image from compressive sensing (CS) based random measurements, but most of them have high time complexity and are inappropriate for patch-based CS capture, because of their serious blocky artifacts in the restoration results. In this paper, we first present a compact network module on the basis of the residual convolution neural network (CNN), which is effective for image reconstruction from non-overlapping patch-based CS random measurements and for blocky artifact removal. Later, we further design an end-to-end network for joint image patch reconstruction and blocky artifact removal, without a separated de-blocky step. By introducing a coding layer into this end-to-end network, we are capable of learning the optimal compressive coding, rather than using Gaussian distribution based random sampling. Experimental results show that our proposed networks outperform the state-of-the-art CS restoration methods with patch-based CS random measurements on synthetic and real data. More importantly, under the learned optimal CS coding, the restoration results could be significantly improved over using traditional random sampling. To demonstrate the effectiveness of our residual CNN based network module in a more general setting, we apply the de-blocky process of our method to JPEG compression artifact removal and achieve outstanding performance as well.
AB - A series of methods have been proposed to restore an image from compressive sensing (CS) based random measurements, but most of them have high time complexity and are inappropriate for patch-based CS capture, because of their serious blocky artifacts in the restoration results. In this paper, we first present a compact network module on the basis of the residual convolution neural network (CNN), which is effective for image reconstruction from non-overlapping patch-based CS random measurements and for blocky artifact removal. Later, we further design an end-to-end network for joint image patch reconstruction and blocky artifact removal, without a separated de-blocky step. By introducing a coding layer into this end-to-end network, we are capable of learning the optimal compressive coding, rather than using Gaussian distribution based random sampling. Experimental results show that our proposed networks outperform the state-of-the-art CS restoration methods with patch-based CS random measurements on synthetic and real data. More importantly, under the learned optimal CS coding, the restoration results could be significantly improved over using traditional random sampling. To demonstrate the effectiveness of our residual CNN based network module in a more general setting, we apply the de-blocky process of our method to JPEG compression artifact removal and achieve outstanding performance as well.
KW - Blocky artifact removal
KW - Compressive sensing
KW - Convolution neural network
KW - Patch-based image restoration
UR - http://www.scopus.com/inward/record.url?scp=85062655830&partnerID=8YFLogxK
U2 - 10.1016/j.neucom.2019.02.036
DO - 10.1016/j.neucom.2019.02.036
M3 - Article
AN - SCOPUS:85062655830
SN - 0925-2312
VL - 340
SP - 145
EP - 157
JO - Neurocomputing
JF - Neurocomputing
ER -