The invention discloses a high-efficiency video coding (HEVC) method based on feature classification. First, standard HEVC is carried out on 10 frames selected from a video sequence, depth features, gradient features and variance features are extracted, and the features are trained offline by using an SVM (Support Vector Machine) to obtain an SVM prediction model. Then, each video sequence is coded, and the first frame is set as an online learning frame. Quick coding is carried out from the second frame, edge detection is carried out, and whether the current CU (Coding Unit) is a CU with complex texture is determined. The current CU is classified as a CU skipping classification in advance, a CU terminating classification in advance or a CU with uncertain classification according to the prediction result. If the current CU is a CU with uncertain classification, the RDcost of the CU is judged. If the RDcost of the current CU is less than the threshold RD_thr under the same depth, theclassification of the CU is terminated in advance, otherwise, a standard rate distortion optimization process continues. The coding complexity is greatly reduced under the condition that the increasein bit rate can be neglected.