STIR transform-based invariant image memory design

Most problems in pattern recognition arise from the translation, rotation and size-changing of the recognised pattern. One solution to this problem is the use of a class of transforms which are simultaneously invariant under shift, rotation and scaling, so called STIR invariant transforms. The algorithms are based on a general transformation where the kernel itself contains the function to be transformed. Thus the invariances are achieved by a kind of self-mapping. Preprocessing of the input signal, such as determination of the centroid or coordinate transformation, is not necessary. The paper gives the results of development work related to the design of an invariant associative image memory based on the use of a STIR invariant transform. The memory is implemented as a software package on a PC and tested with recognition of a large set of human face photographs, which are arbitrary shifted, rotated, scaled, partly covered and corrupted by noise.