Abstract: To solve the problem that traditional Retinex cannot work well in large dynamic range smoke image enhancement, the reasons was analyzed and a new Retinex algorithm was proposed with self- adaptive grayscale stretching. A mathematical model was built to estimate the gray level range of smoke area by calculating local dynamic range and information entropy. By stretching the gray level range calculated and processing the image with Retinex of different scales, the enhanced image was got. Experiment shows that the method can increase the information entropy of large dynamic range image and enhance the contrast of smoke area.