Video Retrieval Algorithm Based On Improved RBF Neural Network

In recent years,with the continuous development of computer and network communication technologies,video has received more and more attention as a medium for information dissemination.How to quickly and accurately find out the video content that people need in massive video data has become the current research focus.Video retrieval technologies currently mainly include text-based methods,content-based methods and deep learning-based methods.Among them,currently the best retrieval effect is based on deep learning methods,the most representative of which are the VGG16 and Alexnet algorithms.However,the neural network algorithms based on deep learning all have the following problems: too many parameters,and overfitting will occur when the training data is limited;the network structure is complex,the calculation complexity is large,and the hardware requirements are too high;The deeper the network structure,the easier the problem of gradient dispersion,which makes the optimization of the model more difficult.In view of the above problems,this paper proposes an image recognition algorithm based on image slices,which uses a multi-level KNN algorithm to construct the center vector,uses the center vector to replace the convolutional neural network as the main feature extraction method,and replaces the multiplication operation with the addition operation.It can effectively reduce the problems caused by excessive neural network parameters and complex structure;secondly,the similarity comparison algorithm of the bag-of-words model is used for image recognition.After the central vector is encoded in this paper,the images and images to be retrieved are constructed according to the encoding The label vector of the image in the library,and then perform the image recognition operation by comparing the label vector of the two.This paper proposes a multi-level KNN-like algorithm in the feature extraction section.The algorithm extracts a slice vector from the feature vector of the image to be retrieved each time,and uses this vector to calculate the Manhattan distance from all the saved center vectors.If a certain center vector is close to the slice vector of the picture to be retrieved,it can be shown that the two are similar in content,and the slice of the picture to be retrieved is replaced with the center vector.If the distance between the slice vectors of the image to be retrieved is greater than the set threshold,indicating that the currently saved slices are not similar to the slice of the image to be retrieved,the slice is automatically added as a new center vector to the saved slice vector set,thereby increasing The coverage of the center vector category.For the obtained center vector set,this paper conducts a comparative experiment with the center vector obtained based on the kmeans clustering method,restores the obtained replaced image,and calculates the Manhattan distance from the original image.It is proved by experiments that the single pixel Manhattan distance obtained by comparing the method proposed in this paper with the original image is 0.0002 less than the absolute value of the distance obtained by the clustering algorithm,and the number of training images required is about50% less than the clustering method.In the similarity comparison part,this paper proposes a label vector comparison algorithm similar tothe bag-of-words model.The algorithm first traverses all center vectors and numbers them in the format of "image name_start row number_start column number_end row number_end column number_channel number_id" so that each center vector slice has The unique id.The image library and all image slices of the image to be retrieved are replaced with the id of the center vector closest to its Manhattan to obtain the label vector of the image,and then the label vector is used for comparison.If the image slice of the image to be retrieved does not belong to any central vector,the slice is automatically saved in the central vector set,so that the system has the ability to learn while training.Finally,in terms of system implementation,this paper introduces the proposed algorithm into the video retrieval system,proves the feasibility of the algorithm,and compares it with the video retrieval technology based on VGG16 and Alexnet network.Under the same platform,the system built on the basis of the algorithm proposed in this paper,the overall running time is 30% less than the current general deep learning-based video retrieval algorithm,and the retrieval accuracy is improved by 1%.Therefore,the algorithm proposed in this paper can lay the foundation for the later development of video retrieval hardware system.