A novel method of Bayesian adaptive estimation was proposed to improve reconstructed slice images based on a photon-counting integral imaging system for three-dimensional (3D) targets in a photon-starved environment. First, a series of photon-counted elemental images were obtained by a photon-counting integral imaging system. Subsequently, based on the Poisson distribution of the photon-counting process, the posterior probability model for photon estimation of the elemental images was established with one local adaptive mean value introduced. The model benefits from the feature of multiple sampling for the same reconstructed voxel by the integral imaging system. Finally, the photon-counted elemental images were updated by calculating the expected value of the posterior probability model and the depth slice images were reconstructed by back-propagating the captured light rays. Experimental results show that the peak signal-to-noise ratio of the depth slice images reconstructed by the proposed method can be 7.4 dB and 8.5 dB higher than that of conventional Bayesian estimation at two scene depths, which greatly improves the quality of 3D target reconstruction.