Research On High-dimensional Deep Feature Compression Algorithm In Soccer Video

Feature compression can reduce the amount of feature data,which is of great significance to alleviate the storage pressure and reduce the consumption of communication during transmission.Feature is the most basic part of neural network and plays a key role in the results of various deep learning tasks.Therefore,reducing the amount of feature data on the premise of ensuring the effect of feature expression has become the key and difficult point of feature compression research.Most of the existing feature compression algorithms tile the feature channels into a large two-dimensional matrix,compress the redundancy in the channels,and rarely consider the redundancy between channels.In addition,the amount of depth feature data extracted in soccer video vision task is large,which brings great computational pressure.In view of the above problems,a high-dimensional depth feature compression algorithm in soccer video is proposed to realize the intra channel compression and inter channel compression of features.The channel refers to the feature map output by each convolution kernel in the deep neural network.Through feature analysis,it is found that there is sparsity in channels and correlation between channels,which indicates that there is redundancy in features and provides a basis for feature compression.For the intra channel,the sparse compression algorithm is proposed.The channels are classified by semantic information.According to the contribution,different channels are treated with different sparseness methods and compressed by intra channel coding.For the inter channel,the sequence correlation compression algorithm is proposed.The feature channels are reordered,so as to increase the similarity between adjacent channels.The feature block matching method is used to encode the motion vector and prediction difference to realize feature compression between channels.Experiments on soccer dataset show that the highest compression ratio of this algorithm is up to 23:1.Compared with the original features,the accuracy of soccer video event classification task using the compressed features is only reduced by about 0.5%.However,this algorithm only compresses the deep features.In the future,all intermediate features can be studied to realize a more general feature compression algorithm.