To reduce the loss of image details caused by traditional median filter, an adaptive median filter algorithm is proposed based on multi-stage noise detection. According to the spatial correlation of pixels, the proposed algorithm detected the types of noise progressively. First, it detected the single noise pixel by counting the number of pixels that had similar values in filter window. Then it extended the filter window in adjacent space to detect the two pixels noise, and added constraints to detect the three or above adjacent pixels noise. Finally, the noise pixel was replaced with the median. In addition, the proposed algorithm can adjust the threshold of pixel spatial correlation discrimination adaptively to meet the processing of different distribution noises. Compared with other median filters, experimental results show that the proposed algorithm can remove noises effectively with maintaining the detail information of the original image.