The invention relates to a hypothesis detection method based on deep learning, and the method comprises the following steps: 1) collecting samples, and creating a hypothesis sample database; 2) constructing and training a yawing preliminary detection model based on the yawing sample database, and fusing low-level local fine features on the basis of high-level semantic features in the feature extraction process of the yawing preliminary detection model; 3) detecting a to-be-detected image by using the yawing preliminary detection model to obtain a preliminary detection result; 4) extracting a mouth area in the preliminary detection result by utilizing the face key point information; And 5) fitting the mouth with an ellipse, respectively corresponding the distance between the upper lip and the lower lip and the distance between the left mouth corner and the right mouth corner corresponding to the long axis and the short axis of the ellipse to obtain a mouth opening degree, and judging whether yawing occurs according to the mouth opening degree. Compared with the prior art, the final detection result is determined by taking the model detection result of feature fusion as the preliminary detection result and combining mouth area feature analysis, and the method has the advantages of high accuracy, high detection completion rate and the like.