Animal’s short-term chewing behaviour is accomplished with discrete jaw movements which occurs through a repeating jaw opening-closing cycle. Recognition of short-term chewing behaviour is the foundation of feeding behaviour analysis and feed/pasture intake estimation for sheep. Several attempts have been made to establish models for short-term chewing behaviour recognition based on the jaw pressure or acoustic signal obtained using wearable sensors. However, such data collection methods have shortages such as difficulty in battery replacement, low stability of the data, and the sensors are vulnerable to damage. A short-term chewing behaviour identification method using computer vision technology was presented, which can extract the frequency and duration of each individual short-term chewing from the feeding video of sheep. Firstly, based on the characteristics of sheep feeding area, a module for target box selection was added in the EfficientDet network architecture. This modified EfficientDet network was employed to detect the three status of sheep mouth, that was opening, stagger of the upper/lower jaw, and closing, in each video frame. Once the sheep mouth status in a video frame was determined, a numerical label was assigned. Then, the regular expression was employed to extract the numerical label sequence segment corresponding to each individual jaw opening-closing cycle. Finally, classification rules were constructed for short-term chewing behaviour identification, where chewing with the side face facing the camera, chewing while facing the camera with the head down, chewing while facing the camera with the head up, and sheep chirping were distinguished. The performance of the sheep mouth status detection obtained by the modified EfficientDet-D0~D4 networks were compared with those obtained by the YOLO v5 and SSD networks. The comparison results indicated that the precision rate, mean average precision, and frame rate of the modified EfficientDet-D1 network was 95.64%, 98.84%, and 28.18 f/s, respectively, which were better than those of YOLO v5 and SSD. Short-term chewing behaviour classification rules, which consisted of EfficientDet-D1 network and regular expression, were applied to the testing videos. The testing results indicated that the frequency and duration of shortterm chewing can be extracted from the videos with the accuracy of 91.42% and 90.85%, respectively. The developed method transformed the video-based sheep short-term behaviour identification problem into the problem of the status label sequences classification, which reduced the complexity of the short-term chewing behaviour classification task. The presented method provided a solution for the automatic short-term chewing behaviour recognition for sheep.