Robustness of regional matching scheme over global matching scheme

Abstract

The paper has established and verified the theory prevailing widely among image and pattern recognition specialists that the bottom-up indirect regional matching process is the more stable and the more robust than the global matching process against concentrated types of noise represented by clutter, outlier or occlusion in the imagery. We have demonstrated this by analyzing the effect of concentrated noise on a typical decision making process of a simplified two candidate voting model where our theorem establishes the lower bounds to a critical breakdown point of election (or decision) result by the bottom-up matching process are greater than the exact bound of the global matching process implying that the former regional process is capable of accommodating a higher level of noise than the latter global process before the result of decision overturns.We present a convincing experimental verification supporting not only the theory by a whiteblack flag recognition problem in the presence of localized noise but also the validity of the conjecture by a facial recognition problem that the theorem remains valid for other decision making processes involving an important dimension-reducing transform such as principal component analysis or a Gabor transform.

An analysis on sub-regional voting scheme

“…For example, the national voting scheme and the regional voting scheme can be applied to facial recognition by 3 considering a face as a nation. A facial image is partitioned into small regions, and the regional voting scheme is applied in the process of matching.…”

“…This demonstrates that there is a certain regional size which provides the best performance of retaining original dominance in the regional voting scheme. If the regional size is selected as 1 x 1, then regional voting would be identical to national voting [3]. Therefore, it is easy to know that the performance of a regional voting scheme is dependent on the size of region.…”

Regional displacement matching scheme for LBP based face recognition.

“…The stability of regional voting was proved by Chen and Tokuda in 2003 [12], and later introduced to the studies of face recognition. To learn this voting scheme, we will first have a glance at the voting problem in general.…”

“…We believe, as supported by [12,14], such approaches lack tolerance for what is called noise (the sudden change of the values of some pixels which makes these pixels no longer corresponding to its original objective), neither do they enhance the ability to tolerant the biological deviations of face features which might happen only in some random regions of the face area and expose the disadvantage of fixed-weight scheme. In fact, a pre-set alignment always draws defects in some cases.…”

Face Recognition Through Regret Minimization.

“…They define stability as invariance to noise, and use two types of noise: uniform and concentrated. They have shown for both binary images [21] and grayscale using the Hamiltonian distance [22] that Regional Voting increases stability to both types of noise. The reason posited for the increase in stability is that Regional Voting is able to contain noise contamination to the regions affected, so that it takes widespread noise to change the classification of the system.…”