Implementing SRCNN in Python using Keras | Image Super Resolution | Tutorial

Hello there, lets go through another great prject, but before we start make sure you know what single image super resolution is. You can check it out here, Link: • What is Single Image S... [Introducion to Single Image Super Resolution]
This video uses Python for performing image super resolution using SRCNN model and is for anyone with a knack for exploring the ever evolving image super resolution field.
Chapters:
00:00 Introduction to SRCNN
03:25 SRCNN Implementation
Resources:
1. Understanding how interpolation works: • Perspective Transforma...
2. Image processing functions & pre-trained model weights: github.com/MarkPrecursor/SRCN...
3. Source Images: mmlab.ie.cuhk.edu.hk/projects/... | Alternate link: drive.google.com/drive/folder...
4. SRCNN Model code: colab.research.google.com/dri...

[@bibhutibaibhavbora8770]

Great video❤ It would be amazing if you make a video on wav2lip architecture. How to customise it andget the best output from the model.

[@AdiTOSH]

Thank you, will look along the suggested lines.

[@loveofmylifesoumyarashmi9972]

hi... thanks for the vdo, bt can I use it for a set of 200 images or more? and one more thing, when both original image and super resolution images are similar ,what is the need of super resolution?

[@AdiTOSH]

Yes, absolutely. You can use it for a set of 200 or more given thats your testing set. For training, we need a much larger set and we are using pre-trained weights for the sake of avoiding high gpu consumption that training typically takes. The SRCNN model is a basic model with just 3 neural network layers. In the world of image super resolution, there are way more deep neural networks with a huge number of layer, way more powerful models that result in astonishingly sharp changes on low resolution images. This model is usually used to learn to fundamentals and understand how image super resolution works. The intention is not to use it to actually perform super resolution on real images captured by our cameras with huge number of pixels. But for less number of pixels, on standard test set, the model performs well even on visual levels. Even, google phones, use image super resolution as post processing to produce picture qualities which are impossible to capture without dslr camera like huge lenses. But the models they use surpasses this basic model that we are studying as a part of research by leagues. However, like any technology, the fundamentals stays the same and without them you cannot comprehend the versions built on top of it. Thus, from study point of view, research point of view, this is an important model, but of little use in actual image super-resolution tasks on pictures captured by cameras. I hope this generates clarity over your concerns, thanks!

[@loveofmylifesoumyarashmi9972]

@@AdiTOSH ooooho ok , now I got it. thank you.

[@user-et3hz2wm9c]

Thank you for great video!! I want to make h5 file by myself so I try to train my model, so I use fit function but it doesn't work. Could you help me?

[@AdiTOSH]

Most welcome! About creating your own h5 file, I am afraid I could not do it either. I observed that google colab is not enough to train SRCNN model and my computer has neither the amount of GPU that training requires. I am afraid I cannot help you here. Thanks.

[@Lachu-hi6gg]

Is there any complete code in google colab for image super resolution using FSRCNN,EDSR,VDSR models.If so please provide the link

[@AdiTOSH]

Hi @Lachu-hi6gg, Sure, here is what I am aware of: EDSR, WDSR, and SRGAN: ◦ The GitHub repository provides implementations of EDSR, WDSR, and SRGAN for single image super-resolution. You can load pre-trained weights and apply super resolution to images 2. ◦ Check out the code and examples in the repository here. github.com/krasserm/super-resolution

[@My-bself]

Bro this is my project incan directly run the cofe in collab or any error comes according to it

[@My-bself]

Any one help me to run the code broo

[@AdiTOSH]

Yeah bro, you should be able to directly run the code, just follow along with the video, it should be fine. You will not get any error.

[@My-bself]

@@AdiTOSH image is not downloading broo

[@My-bself]

That's the problem

[@AdiTOSH]

@@My-bself I see, that should not be happening. It does download for me, can you try this link in that case, drive.google.com/drive/folders/1e5u38BG6lBjUm24uRb4WS3lmms0r6UtM?usp=share_link

[@saurabhsingh5948]

can you please put a SRCNN with a training data rather than pretrained model weights

[@AdiTOSH]

Hi @saurabhsingh5948, I see that you are looking to creating your own h5 file rather than using one. I am afraid but I could not do it either. I observed that google colab is not enough to train SRCNN model and my computer has neither the amount of GPU that training requires. Thank you.

[@ramakrishnamiryala]

The source images link is not working.

[@AdiTOSH]

Hi @ramakrishnamiryala, The link does not open a webpage but instead downloads the zip directly through your browser. Can you please check if your browser if blocking the download or try a different browser. Thanks.

[@lucc8703]

P r o m o s m