Research On Deep Hash For Video Copy Retrieval

In recent years,a lot of video is expanding at an amazing speed.How to enable users to quickly find the target video in a huge amount of data has become one of the topics that have drawn much attention at present.For example,the user wants to retrieve the full version of the video by using a slice of sequence,where the main difficulty is how to quickly and accurately retrieve the copies of videos which have same contents but have different bit rates,resolution,frame rate,encoding,etc.With the development of image processing technology,people tend to search videos based on contents rather than keywords of manual annotation.What's more,copy video is more suitable for content based retrieval when the user cannot generalize the semantics.Mapping video features to hash codes can greatly improve the retrieval speed.Therefore,how to design a valid and accurate hash code for video retrieval is the key point of this paper.There have been two main types of research on video hash retrieval recently,retrieving video from images hash and video hash over video.To retrieve video through image hash matching only needs to consider the features of the space of the image.The feature extraction is relatively simple while the retrieval result is not accurate enough.The video hash over video needs to consider the features of the video space and time.The retrieval result is more accurate while the feature extraction is more difficult.In order to improve the retrieval precision,this article uses the method of video hash over video.Therefore,this paper proposes a temporal-spatial fusion hash code.For the spatial hash,traditional video retrieval often uses SIFT based local features.In recent years,with the development of the neural network in image feature extraction,neural network has been better than the traditional method.So this paper uses neural network to extract deep video hash;For the temporal hash,the polymerization method based on frame difference not only absorbs the time information but also reduces the feature dimension.Fast video retrieval by spatial-temporal hash with multiple hash table indexes and two pruning methods is then to verify the accuracy and efficiency.The main innovations of this video hash are as follows:In the spatial domain,this paper presents a new scalable compact deep hash fingerprint whose innovations are in the below.1)Features of the very last pooling layer instead of the classification layer and fully connected layer;2)Aggregation of the Deep Learning features that provide compact and scalable content representation.3)Hash generated by direct binarizing the Fisher Vector with component/dimensionality priority scalability.In the temporal domain,this paper proposes a hash fingerprint of polymerization aggregated discrete cosine transform(DCT)coefficients.Innovation is that frame difference based DCT with Gaussian filter and hash fingerprint of the top energy concentration of DCT.Through testing the public data sets TRECVID and CC_WEB_VIDEO,the advantage of this method is that the hash code PR curve designed by this paper is superior to other hash main methods for the accuracy of retrieval.In view of the retrieval speed,just 5 seconds of short video query can automatically retrieve the video segment,the average retrieval time of 4ms,false positive rate of 0.3%,faster than TCSIFT.