In order to avoid running a classifier at every possible position on the input image (when most won't contain an object), we can use some external method to propose likely regions to us. One method that can do this is called Selective Search.

The region-proposal method will provide blob-like rectangular regions in the image, which are likely to contain objects of interest. These regions are the candidate areas for the presence of an object of interest. A CNN classifier is then applied only on these proposed regions. This cuts down massively on the number of crops sent to the CNN for classification, as compared to the sliding window approach.

This particular approach was suggested in 2013 and was known as R-CNN: Regions with CNN. The following diagram depicts the process of R-CNN: