I am a complete beginner in image analysis and I would like some guidance on this problem. I need an algorithm that can take a query logo and match it to the most similar one in a database (with thousands of entries). Since an image of a logo can have many slight differences from another image of the same logo, the algorithm would need to be tolerant to this. For example, consider these two Huawei logos. They are of course the same logo but on a detailed level they have differences:
- Font
- Shading/Color Hue of the flower
- Exact positioning of the shapes
I have tried both pHash and SIFT but both seemed to return very unsatisfactory results. In both cases, I preprocessed the images to trim out extra background space, have the same size, and quantify the color spectrum.
My next idea was to perform a euclidean distance similarity or cosine similarity on the pixel information directly. Though this would have the issue that the shapes in two images representing the same logo will not necessarily overlap.
Does anybody have any suggestions of how this can reasonably be done?
