In this paper a new efficient method for detecting the impulse noise from the corrupted image using extreme learning machine (ELM) is proposed. An improved version of the standard median filter is suggested to remove the detected noisy pixel. The performance of proposed detector is evaluated using classification accuracy. The results show that our detector is robust even at higher noise density. Results illustrate that proposed filter provides better performance in terms of PSNR than many other median filter variants for Salt and pepper noise. . The suggested technique yields significantly good results both in objective and subjective judgments of image quality.