An AI-based approach to auto-analyzing historical handwritten business documents: As applied to the Kanebo database

Matching salient points is a key step in visual tasks. However, many of the existing feature representation methods that are widely applied to these tasks, such as scale invariant feature transform (SIFT), suffer from a lack of representation invariance. This shortcoming limits the image representation stability and salient-point matching performance, particularly when images with a great deal of noise information are being processed (e.g., historical documents). We propose a general and effective transformation approach called RIFT (reversal-invariant feature transformation) for feature-robust representation. RIFT achieves gradient binning invariance for feature extraction by transforming the conventional gradient into a polar one. Experimental results on the Kanebo database and three fine-grained reference classification datasets demonstrated that RIFT can robustly improve the performance of local descriptors for image classification without sacrificing computational efficiency.