In this paper, a double-threshold filter strategy was proposed and experimentally tested to improve the accuracy of photon counting imaging with an electron multiplying charge-coupled device, by taking into account the addition bias noise and Pseudo-photon noise during photon counting mode. Photon counting event detection was achieved by using amplitude threshold and frequency threshold filter to eliminate the bias noise and Pseudo-photon noise. Based on Bayesian estimation on multi-imaging, an optimal detection rule was found with minimum error cost. Experiment indicates that double-threshold filter strategy improves the photon detection accuracy than usual integrate method, and is suitable for imaging in ultra-low-light conditions such as biological dim light detection.