Near-duplicate Video Retrieval Based On Deep Autoencoder

As an important information carrier in the Internet era,the video data play an increasingly important role in our daily life.Recently,mass near duplicate videos appeared on the Internet and redundant videos have brought inconvenience for video management.How to detect these near duplicate videos quickly and accurately has becomes a hot research topic.Content-based near duplicate video retrieval consists of three parts:key frame extraction,feature extraction and feature matching.In the feature extraction part,the current methods mostly use the visual features or the local features of the images,the dimensionality of these features is usually quite high,and the representation ability of them is limited,which leads to weak retrieval performance.In recent years,major breakthrough has been gained in the field of deep learning.Deep learning could achieve better results in many problems than traditional methods.In the field of image processing,deep neural network can automatically extract more abstract features of the image,and then use these features to complete the image recognition or classification.However,the method based on deep neural network has not been widely used in the field of video retrieval.Based on this,this thesis has done the following research work:1.An adaptive key frame extraction algorithm based on shots is proposed.The method first divides the image into many blocks and obtains the color histogram of the image,then calculates the difference between two key frames by weighted summing.Next,it calculates the difference between the current frame and the mean value of the interframe difference in a frame set,and compares the difference with a threshold to determine whether the current frame is a boundary frame of a shot.The experimental results show that the method can extract the key frame better.2.A near duplicate video retrieval method based on deep autoencoder is proposed.This method uses the deep auto-encoder neural network to extract the features from key frames.In order to compare the effect of this method with the methods based on other image features,experiments were performed on standard near duplicate video retrieval data sets.The experimental results show that the proposed method is much better than the retrieval methods based on image color feature and LBP feature.