Abstract: Abstract: Since cattle facial information is very rich, ranging from skin color, chewing conditions to health status, it is of great importance for cattle disease monitoring. This paper proposed a scheme on cattle facial automatic extraction to cope with the problems caused by different camera angles, varied illumination and partial occlusion. Three classical human facial contour extraction algorithms were employed in the study: Supervised descent method (SDM), local binary features (LBF) and fast active appearance model (FAAM). For contour extraction performance evaluation, 800 cattle facial images were collected from a cattle farm at Northwest A&F University. For each cattle face, 29 facial feature points were manually labeled. In order to cope with cattle face scale and rotation, all the labeled facial feature points were aligned to a normalized model. Compared with human face, cattle face height-width ratio was much larger. Based on the characteristics of cattle face, this study used the AdaBoost detector to train cattle face detector. Because cattle face was rather long, we cut the image to the size of 15×25 pixels. And the background was a negative sample with the same size. The detection rate of the cattle face detector for 800 cattle face images was 96%. Considering the characteristics of cattle face, we improved the algorithms and optimized our parameters. Also because of the long face, a split model was used to initialize the cattle face model. The first part included an eye contour with cheeks on both sides, and the second part included the nose and mouth contour. The results showed that the accuracy of the contour extraction was improved significantly. We then analyzed and compared the time efficiency and the accuracy of the 3 algorithms. Finally, the performance of each method was evaluated by the Euclidean errors normalized by the left and right corners of the eyes and their corresponding computational time costs. The average computational costs of the 3 contour information extraction methods were 1.75, 0.35, and 60.62 s respectively. The average pairwise Euclidean errors normalized by the left and right corners of the eyes were 0.0188, 0.0245, and 0.0184 pixel. The experiment verified the feasibility and practicability of the facial contour extraction methods. Results showed that the FAAM algorithm achieved the highest accuracy with the minimal alignment errors while the LBF algorithm was the most efficient. Therefore, in the process of facial contour extraction, we can choose proper algorithm in varied situation which requires different accuracy and efficiency. The contour extraction algorithm can effectively extract cattle facial contour information, which provides a theoretical basis for the further analysis of cattle facial expressions. As this is the first time that the contour extraction algorithm is applied to cattle face analysis, the present study serves as a good guide for other researches, which provides feasible data for computer vision based cattle disease analysis and a comprehensive guideline on cattle facial analysis under different circumstances in intelligent farm.