The photon counting image is scanned by multi-pixel photon counting detector point by point under the environment of 10-4 lx according to the principle of photon counting. To present more details and get a high definition image, the Bayes-Shrink threshold and the improved new symbol functions are used to realize image de-noising preprocessing first. Then, in the stage of image reconstruction, the low-frequency coefficients are set to zero to reconstruct the image with high-frequency coefficients after processing and it is set as a virtual channel to make the number of observation signals equal to the number of signal sources. Finally , the fast independent component analysis noiseless separation model is used to separate the photon counting image from noise by blind source separation. The experimental results show that the peak signal to noise ratios of the image are improved by 16.39%, 10.18%, 5.20%, respectively, compared with the soft, hard and new symbol function de-noising algorithm. The image after removing noise is also good to protect the edge details, and the visual effect is good.