摘要： The scene change detection (SCD) is the necessary issue for further video applications. Current SCD techniques employ the Convolutional Neural Networks (CNN) for image feature learning and outperform traditional optical ?ow methods. However, those methods use traditional machine learning methods for scene classi?cation. In the paper, we present a novel framework through deep learning networks and image matching method. We employ the ResNet for video image training and learning to classify the video frames into different categories. The classi?ed frames can be used for scene change detection. For those predicted scene changed frame pairs, we adopt the SIFT algorithm to extract features of them and through image matching to remove the failure detection. We perform the experiments on several different kinds of videos. The proposed framework can detect the scene changing parts and divide the images into scene units with low error rate.