Conclusions and Future Work

 

Having achieved the main objective of the work, some conclusions taken from it and future work to be developed are presented.

 

Conclusions

 

The implemented system achieved the objectives proposed: a hardware co-processor which implements the iterative hard thresholding algorithm. The system was developed to recover 8x8 patches from 16 measurements and assuming a sparsity of 8. Furthermore, the chosen sparsity domain was the discrete cosine transform.

 

The final implementation showed the following results:

 

• Image quality results for a 4 megapixel recovery, depending on superimposition level :

  • PSNR : between 12.7dB and 19.4dB

  • SSIM : between 0.09 and 0.42

• Maximum frequency : 100MHz

• Power consumption : between 2.1 and 2.25 Watts for frequencies from 50MHz up to 100MHz

• Speedups of 190 times using a 100MHz implementation compared to a MATLAB implementation running on a 3.4GHz CPU. If the frequency is taken into account, the speedup is 6460

 

Future Work

 

A future work over the presented results of this project could focus on:

• Implementing several recovery processors on the programmable logic to fully utilize the parallel capabilities of the system

• Changing the sparsity transform used to evaluate the difference in the quality of the recovered image and run time

• Changing the specifications on the level of the patch size, number of measurements and sparsity level

• Trying new approaches for the step size computation