Performance bench for CNN 0 branch

Condition

• Tested on AARCH64 ( Rockchip RK3399, Custom kernel 4.4.179-perf )
• "cnn-0" is a different version of zvezdochiot.

Version 0.1.4.18 - fetched

> bin/srcnn Pictures/butterfly.png
srcnn : Super-Resolution with deep Convolutional Neural Networks
(C)2018 Raphael Kim, (C)2014 Wang Shu., version 0.1.4.18
Built with OpenCV version 2.4.9.1

- Scale multiply ratio : 2.00
- Image load : Pictures/butterfly.png
- Image converting to Y-Cr-Cb : Ok.
- Splitting channels : Ok.
- Resizing splitted channels with bicublic interpolation : Ok.
- Processing convolutional layer I ... completed.
- Processing convolutional layer II ... completed.
- Processing convolutional layer III ... completed.
- Merging images : Ok.
- Converting channel to BGR : Ok.
- Writing result to Pictures/butterfly_resized.png : Ok.
- Performace : 21305 ms took.

Version 0.1.5.21 - cnn-0

> bin/srcnn Pictures/butterfly.png
srcnn : Super-Resolution with deep Convolutional Neural Networks
(C)2018 Raphael Kim, (C)2014 Wang Shu., version 0.1.5.21
Built with OpenCV version 2.4.9.1

- Scale multiply ratio : 2.00
- Image load : Pictures/butterfly.png
- Image converting to Y-Cr-Cb : Ok.
- Splitting channels : Ok.
- Resizing splitted channels with bicublic interpolation : Ok.
- Processing convolutional layer I + II ... completed.
- Processing convolutional layer III ... completed.
- Merging images : Ok.
- Converting channel to BGR : Ok.
- Writing result to Pictures/butterfly_resized.png : Ok.
- Performace : 39749 ms took.

Conclusion

• Idea for Convolution I and II in a time was good. but where is performance ?
• Which is better ? Losing performance but saves memory in time for over 10 seconds ? Nonsense.