This thesis deals with content-based image retrieval from large-scale image databases. The aim is to nd the original version of a modied image. The modications are suppressed by normalization, the search itself is based on image intensities. Candidates for match are retrieved from our binary search tree with use of stochastic backtracking. The candidates proceed to one-by-one comparison and ltered to correct matches. The system has been implemented and is used in practice. Our test made on database of 25 milion images prove that the retrieval is fast, the error ratio is low and the success rate is good for practice, varying for dierent degradations.