The mounting behavior of pigs is generally manifested as a pig puts two front legs on the body or head of another pig which stays lying or dodged quickly. Mounting between pigs often causes epidermal wounds and even fractures, which reduces animal welfare and affects the economic benefits. Therefore, it is necessary to isolate the mounting pigs in time. In view of the low degree of automation of current mounting behavior detection of pigs, an algorithm based on Mask R-CNN was proposed to recognize the mounting behavior of pigs. Firstly, the top view videos of pigs were shot, and the dataset labels were made by Labelme. The transfer learning was applied to train the ResNet-FPN network to obtain the pig segmentation result and extract the mask pixel area in each sample. The value of the minimum mask pixel area in each sample was extracted in order to build an empirical sample set for mounting behavior recognition, and the discriminant threshold of the mounting behavior of pigs was determined. In the experiment, the test dataset was used to evaluate the pig segmentation network model and the mounting behavior recognition algorithm. The segmentation accuracy of the network was 94%, and the accuracy of the mounting behavior recognition algorithm was 94.5%. The experimental results showed that the algorithm can effectively detect the mounting behavior of pigs and provide support for livestock breeding automation.