I'm looking for a method of detecting multiple, arbitrary, possibly overlapping objects in a dataset.
An example image is shown below. I would like to find an algorithm to detect that there are 2 principle characters in the image, their approximate offsets, and output the image corresponding to each character/icon.
I believe there are common methods for detecting characters after convolution kernels have been trained, but this needs to work on arbitrary characters/icons etc.
I believe a SVD/PCA analysis could easily extract common/repeating elements from rows/cols of the image, or from nicely framed segments of the image (eigenfaces), but i am unaware of how to go about this problem where the icons are local, but in arbitrary locations, and untrained.
I imagine that there is ~some~ mathematical algorithm or set of algorithms, similar to an SVD, which allows such detection and classification generically and mathematically, without eg. ad-hoc segmentation based on background color.
Questions
- Is this a complex/trivial problem?
- What search terms should I use to get closer to a solution?
