A Scheme for Linear Recognition Based on Self-Determined Intercategory Features

摘要

In a pattern recognition problem, where observations are described by a mxn grid, it is often the case that each pattern is reduced to a “mask” and discrimination is performed by comparing an unknown to each mask. Assignment is then determined by the closest mask. This paper discusses a preprocessing technique for feature extraction to reduce the size of the masks or to extract those mask sub-areas most pertinent for recognition. Statistical tests are applied to determine uncommon regions (i.e., differences) between masks. All subsequent recognition is based on and emphasizes these uncommon regions (as distinguishing features). The resulting weights of this method are controlled by differences between the groups and thus cannot be separated into a characteristic set of weights from each individual group. Moreover, this method provides the freedom to select the level of closeness that the categories must satisfy.