Research On Video Object Retrieval Technology Based On Deep Learning

Usually,a traditional video retrieval system only allows using keywords as its input,and it retrieves the relevant video through the artificial labels.As the amount of video data increasing,the cost of manually tagging video content is increasing,and the retrieval results are unable to meet the daily needs of users.Content-based video retrieval uses images or videos as the input to the system,and extracts the content of input as retrieval target.The efficiency of system could be greatly improved,because it dosen't often need human participation.The main work and innovation of this master thesis are as follows:(1)In the video structuring process,this paper studied the problem of the traditional key frame extraction algorithm that the number of key frame is difficult to determine.Then,it proposes a method to calculate an adaptive threshold using the mean and standard deviation of the inter-frame difference.It could find the changes of video content by counting the number of inter-frame difference greater than the threshold.Finally,on this basis,it obtains and extracts the number of key frames.The result of experiments show the validity and generality of this method as the number of key frames obtained by our method is consistent with the optimal clustering number of image frames.(2)For the problem that the results of traditional k-medoids clustering algorithm are susceptible to the initial clustering centers,this paper proposes an improved clustering keyframe extraction algorithm.It uses SOM clustering algorithm to cluster image frames initially.Because the results may not be optimal,we use k-mediods algorithm to optimize the results,which further solves the impact of isolated frame.Compared with the traditional key frame extraction algorithm,the keyframe extraction algorithm in this paper can balance the fidelity and compression ratio well and achieve excellent clustering results.(3)For the lack of semantic information in traditional image features,this paper proposes a deep feature extraction algorithm based on feature points.The algorithm extracts local convolution features of the images by pre-trained CNN,and then SURF or ORB points to extract the critical regions of the image where the local convolution features are more discriminating.Finally,it aggregates the subset of local convolution features into a compact feature descriptor.The result of experiments show the accuracy of this deep feature is better than other traditional artificial feature.Compared with other selective convolution feature algorithm,the proposed method effectively reduces the time of feature extraction.(4)In order to realize video object retrieval based on deep features,we uses MATLABGUI tool to integrate the key frame extraction and feature extraction algorithms proposed in this paper,and realizes a simple and practical video object retrieval system.The system uses image to retrieve video clips which have similar content,realizes the functions of video shot detection,key frame extraction and feature extraction and retrieval.Finally,this paper verifies the usability of the system through experiments.