Attitude estimation of satellite targets with the Inverse Synthetic Aperture Radar(ISAR) image sequences is a significant but challenging task. Existing estimation methods are normally focused on the extraction of critical corners or linear components from the image, which are hard to meet the real-time requirement and insufficient to exploit the prior of imaging characteristics. This paper presents a method for estimating the attitude of satellite targets based on imaging characteristics and regression networks. The imaging characteristics of satellite targets under various attitudes are firstly determined in advance and serve as the theoretical basis for subsequent dataset annotation. Thus, different from the traditional classification problem, a regression network and an estimation framework suited for attitude estimation are established. Finally, electro-magnetic simulation in millimeter frequency is carried out to validate that the proposed method can control the average attitude estimation error within 3.5°.