2022 IEEE 8th International Conference on Smart Instrumentation, Measurement and Applications (ICSIMA)

Deep neural networks have been used extensively

for inverse problems in image processing research. Subsequently

the interest has been shown to theoretically model deep nets

with celebrated sparse coding theory resulting in emergence

of convolutional sparse coding (CSC) theory. The CSC theory

which is a special case of sparse coding works on the premise

of representing underlying data (natural or biomedical images)

with learned fiters/dictionaries and their corresponding sparse

feature maps. The dictionaries have special structure unlike their

counter part of sparse coding and pursuit algorithms works

on global scale instead of patches. This global pursuit results

in mitigating the effects of patch aggregation process during

traditional regularization-based techniques. Further extending

the CSC model, the learning features are again processed

through the CSC model representing them with another layer

of fiters/dictionaries and their corresponding sparse maps. This

process is continued until the last layers of model making

the multi-layer convolutional sparse coding model. The pursuit

algorithms can be employed layer wise of a global pursuit

settings utilizing iterative thresholding algorithms. In this work

we have implemented a ML-CSC model for the reconstruction of

Knee MR images on different CS ratios. The results have been

compared with state of the art ISTA-Net+ model in terms of

PSNR/SSIM and restoration times for different CS ratios. It is

shown that ML-CSC has better quality results than the state of

the art ISTA-Net+ model.