Two ideas, depending on requirements.
1. Monte Carlo Markov Chain: model the strikes as linear sections, with free parameters given by its length, height, width etc. Begin search at local peaks. Constrain model parameters until suitable fit found.
Pros: robust.
Cons: could be slow
2. Cross-correlate with a set of rectangle templates. Accept above a threshold.
Pros: faster
Cons: trial and error to get threshold.