Analysis of Profiling Data

All the Similarity metrics were calculated for an image of size 4000x4000x3 MSE, PSNR, SSIM, MS-SSIM, UQI are calculated. Profiling data has been committed to differ.

Analysis of Profiling Data - some important points

Total elapsed time = 172.99 seconds The following 5 main functions responsible for 81% of total time

calcMSSSIM::compare(_IplImage*, _IplImage*, Colorspace) [4] = 87.79sec

calcSSIM::compare(_IplImage*, _IplImage*, Colorspace) [3] = 17.35sec

calcQualityIndex::compare(_IplImage*, _IplImage*, Colorspace) [17] = 15.78sec

calcMSE::compare(_IplImage*, _IplImage*, Colorspace) [43] = 2.66sec

calcPSNR::compare(_IplImage*, _IplImage*, Colorspace) [44] = 2.66sec

Calculatin SSIM takes most of the time i.e. 65% since it is used by both SSIM and MS-SSIM

calcSSIM::compare(_IplImage*, _IplImage*, Colorspace) [3] = 65.8% of total time spent = 104.08sec

Individual functions in order of most time expensive - % time of the total time elapsed

cvSmooth = 31.9% 50.51 secs 211 calls

cvPow [11] = 11.6% 18.34secs 180calls

cvMul [14] = 10.8% 17.11secs 168calls

cvDiv [22] = 6.7% 10.59secs 78calls

cvSub [27] = 5.2% 8.20secs 126calls

cvAdd [34] = 3.5% 5.46secs 84calls

cvConvertScale [33] = 3.5% 5.48secs 187calls

cvAddS [38] = 1.7% 2.70secs 144calls

cvCvtColor [41] = 1.7% 2.66secs 38calls

sum total of above = 76.6%

With this anlaysis we can observe and conclude that we can improve the performance of the Similarity Metics Algorithms if we make the about functions fast.

Next task is to use OpenCL and CUDA to improve efficiency.