1

I've researched a lot and could not find a definitive answer. What kind of image color is most used for pHash input to generate the hash/fingerprint.

For example I have a target image that I'm looking for within a source image, but the target can have many colors and shades, but the shape is always the same (ex: tulips). I have experimented with the image as is, turned gray scale and threshold (pure black and white). I know most pHash libraries will gray scale the input first before the hash is made.

But before I move forward is pre-processing the image color worthwhile? (ignoring size and rotation, and assuming source and target are the same for both)

P_C
  • 185
  • 2
  • 11

1 Answers1

2

So after testing and more research it's best to use the original colored image. Most pHash will gray scale an image regardless, so performing a gay scale followed by the internal gray scale actually produced poor results. The same goes for Threshold (pure black and white). There were more collisions and many more false positives.

I used a 64 bit pHash and worked very well. I also tried with Wavelet Hash which was good for color changes but not good for overall matching.

What worked for me is a large data set that was feed into a BinaryTree. This way the look ups were fast and had many examples to compare to. For Java I used: https://github.com/KilianB/JImageHash

P_C
  • 185
  • 2
  • 11