This paper presents an effective noise reduction method for both RTS noise and photon shot noise based on a multi-camera configuration composed of multi-aperture optics and a semiphoton-counting-level CMOS image sensor. We model noise characteristics of the CMOS image sensor with a probability distribution based on measured read noise including RTS noise and shot noise. We estimate the average number of photons by the maximum likelihood estimation with all corresponding pixels in the multi-aperture optics. We compared the noise reduction performance of several schemes when the incident photon number was assumed to be 2 for each aperture. 2 out of 9 apertures showed RTS noise. Simple averaging gave effective noise of 0.68 e^-<small>RMS</small>, where RTS noise still exist. Selective averaging, which minimizes synthetic sensor noise, gave that of 0.60 e^-<small>RMS</small>. Although RTS noise was removed, photon shot noise was less suppressed because only 7 apertures were considered in averaging. On the other hand, with the maximum likelihood estimation, the effective noise became 0.48 e^-<small>RMS</small>, and both RTS noise and photon shot noise were reduced.