The existence of blind pixels significantly affects the imaging quality of infrared cameras. Scene-based methods for blind pixel detection and compensation can effectively solve such problems. This study proposes an improved local "3?" method. The average noise of the infrared image is obtained by calculating the three-dimensional noise of the image; according to this average noise, the minimum criterion for blind pixel detection is obtained. Subsequently, real-time dynamic detection and compensation for blind pixel is performed by local "3?" and median filtering method; furthermore, the method was applied to a self-developed mid-wave infrared camera. The results of the blackbody imaging experiment show that, compared with the radiation calibration method, the coincidence degree of blind pixel detection of our method can exceed 82% on average. The proposed method has the same effect of detection and compensation for blind pixel compared with the traditional local "3?" method; however, it can reduce the over-detection rate of blind pixel by more than 30%.The results of imaging experiments of ground scene show that the proposed method can effectively restrain the blind pixel, and no apparent abnormal black and white spots were found in both daytime and nighttime images captured by the infrared camera. Additionally, the scene details in the images are rich, and the image quality is excellent. Therefore, the method presented in this study has good performance in real-time dynamic blind pixel detection and compensation. It is feasible and effective for use in self-developed mid-wave infrared cameras.