Hi All,
I am using variational autoencoders as machine learning model for UI layout generation. Those UI layouts are given as images (matrices) to the machine learning model as input.
The problem is after the model learn the data and it is supposed to generate new examples, the data that is generated is a bit continous.
So my professor suggested to apply Quantize pooling as a post processing transformation to get discontinous images.
It helps also to remove the blurring or the closely graduated pixels in the image.
Anyone can help me with a useful link or paper on how this Quantization works on images ?
I will implement it later using Python.
Can you check the following link:
https://arxiv.org/pdf/1510.09041.pdf
https://pdfs.semanticscholar.org/25f1/b0fe0526e69c8c0d46b102af1d78ca011096.pdf
https://medium.com/comet-app/deep-learning-model-compression-for-image-analysis-methods-and-architectures-398f82b0c06f
http://openaccess.thecvf.com/content_cvpr_2018/papers/Sun_Feature_Quantization_for_CVPR_2018_paper.pdf
Quantization is used in JPEG and JPEG2000
You can also look into gray level quantization
I suggest to add some generated images with your question that someone can look and easily suggest solution.
Maybe there are many solutions other then Quantization. Beacuse as far as i know, Quantization is just normalization of pixel values between some range.
Article Integrating salient colors with rotational invariant texture...
Thank you for your suggestions. I tried applying quantization, sampling and filtering as well.
I can say that I got slightly better results with quantization using K means .
- Badges
- Science topic
- Similar topics
- Computer Science
- Images