An implementation of the image matching algorithm described in An Image Signature For Any Kind Of Image by H. Chi Wong, Marshall Bern, and David Goldberg . The algorithm is designed to detect nearly identical images, not images with the same conceptual content.
By default, the library offers two primary functions: get_buffer_signature(rgba, width) and cosine_similarity(a, b).
The former takes a pre-processed slice of u8s with each chunk of four representing the 8-bit red, green, blue, and
alpha of a pixel, the latter two result vectors to compute their similarity. Per the source paper and our experiments
in this research images with a similarity greater than 0.6 can
be considered likely matches. If the tuning methods described below are used, additional research will likely be needed
to assess a new cutoff.
If the img feature is used, also provided are get_image_signature(image) and get_file_signature(path) which use
the image library to handle unpacking the image into an rgba buffer. All signature
functions also expose tuned versions which allow tweaking the crop percentage used during the signature computation,
the size of the collection grid which controls the length of the feature vector produced, and the size of the
square around each grid point averaged to produce a value for that point. It's recommended to study the algorithm
closely before embarking on tuning, as the effects of these nobs are not immediately obvious.
- Additional unit testing.
- Experimenting with parameter choice is underway in this repo.
- Experiment with widening the possible values of each dimension in the produced signature. Presently per the paper they
are all integers in
[-2, 2]. It will likely require experimentation around a new suggested vector similarity cutoff.