Abstract: The watershed algorithm leads to over-segment when it was used to segment an infrared image. An adjustable bias-anisotropic diffusion filter based on CNN-PDE for smoothing an infrared image was studied. In order to remove the residual noise which can not be smoothed by filter, the smoothing coefficient and constrain coefficient were used to threshold the gradient image. The result of watershed segmentation of a practical infrared image shows the presented method can restrain over-segmentation effectively.