For dither video sequences in the tracking, detecting, combating situations, the observed target blurred, which is not conducive to the dynamics of the tracked target identification. This paper presents a semi-supervised learning mutual information entropy PN image stabilization algorithm to enhance the ability to jitter frequency video sequences, and reduce motion estimation error rate. PN semi-supervised learning of mutual information image stabilization is based on mutual information entropy theory for image motion estimation, on this basis, build parameters and jitter parameters is observed for motion compensation loop structure. Through the motion compensation parameters, semi-supervised PN learning are observed and performed. Mutual information entropy key parameters are corrected to ensure motion-compensated inter-frame parameter estimates and the actual amount of image jitter error is minimized and the output stability. By using mutual information entropy P-N learning semi-supervised learning algorithm for mutual information entropy image library jitter test comparison, motion estimation error rate reduced to 1%, to further enhance the ability of video sequences de-jitter frequency.