In practical biometric verification applications, we expect to observe a large variability of biometric data and single classifiers may not be very accurate. In such cases, fusion of multiple classifiers may improve accuracy. Statistical dependence of classifiers has recently been shown to improve accuracy over statistically independent classifiers. In this paper, we focus on the verification application and theoretically analyze the OR fusion rule to find the favorable and unfavorable conditional dependence between classifiers. Favorably dependent correlation filter based classifiers for the OR rule are designed on the fingerprint NIST 24 plastic distortion and rotation datasets. For the plastic distortion dataset, unconstrained optimal tradeoff (UOTF) correlation filters were used because of their distortion tolerance and discrimination capability; and for the rotation dataset, optimal trade-off circular harmonic function (OTCHF) filters were used because of their tolerance to geometric rotation. On the plastic distortion dataset, three favorably dependent classifiers were designed on different distortions of the finger, each with an EER of 15.7%, 14.3%, and 9.8%
respectively. The OR fusion of these three classifiers has an Equal Error Rate (EER) of 1.8% while the best single UOTF based classifier has an EER of 2.8%. On the rotation dataset, five OTCHF filter based classifiers were designed for tolerance to different rotation angle ranges of a finger with an average individual EER of 38.8%. The OR rule fusion has an EER of 14.6%; whereas the best single OTCHF filter has an EER of 27.7%. It is also shown that the best fusion rule is the OR rule for these classifiers that were designed to be favorable for the OR rule.