The infrared image background is obtained by morphological filtering, and the infrared images are nomalized using the estimated noise mean and variance in a local window. Composite frame is formed by normalized image dara, and target detection is performed on the composite frame. Thus, target and noise statistics in a composite frame are obtained. Considering the different distribution of noise and target intensity, the projection algorithm is further modifed to impove its signal to noise ratio. Simulations of real infrared images proved that the modified projection algorithm can successfully detect targets with SNR equaling 2.