An Improved VLAD Algorithm Combining Attention And Non-local Description Features

Obtaining important information by encoding image features is an important field of artificial intelligence.For better image description information,feature encoding methods based on image processing emerge in endlessly,whether it is the direction of traditional image processing or the direction of deep learning.It is widely used in image classification,target retrieval,and motion recognition based on video streams.VLAD?Vector of Locally Aggregated?Descriptors)have developed research in the fields of traditional image processing and deep learning.However,there are three shortcomings.1.The VLAD algorithm based on traditional image processing can output local aggregate descriptive features for the characteristic of image data,but the features lack the process of learning optimization.The output features and the target label Errors cannot be lowered along the error direction.2.The VLAD algorithm based on deep learning solves the problem of VLAD's optimization of image feature learning,but because the clustering center is initialized and adjusted through network training,if the key areas of image features are strengthened,it will be better divided into clustering center area,output integrated local aggregation feature vectors;3.Due to the final output of local aggregation feature vectors by VLAD,useful information will be lost to a certain extent,affecting the network model identification.In response to this,this paper proposes an improved method of VLAD algorithm that mixes attention and non-local description features.The Attention-NetVLAD network structure based on the attention mechanism is mixed with a non-local description feature based on the correlation between features.The attention mechanism strengthens the image features and the image characteristics.At the same time,the non-local description feature calculates the correlation between the local features,Supplementing the non-local relevant information of the image data,perfecting the output image feature description,and the feature description is more discriminative.The innovations are as follows:?1?Attention-NetVLAD based on attention mechanism is proposed.By using convolution operation to replace the traditional VLAD hard distribution,it is converted into soft distribution and becomes a part of the structure of the neural network.Features.Attention-NetVLAD is divided into two parts,"Space-Attention-NetVLAD based on spatial dimension"and"Channel-Attention-NetVLAD based on channel dimension".The attention-enhanced feature map is input into the VLAD,the residuals of the feature points and the cluster center are calculated,the soft distribution weights are calculated,and finally the residual weighting is performed to output the local aggregated feature vector.Since the key areas of the feature map after attention enhancement are weighted,the VLAD part is easier to divide the cluster center area according to the key areas of the feature map,which promotes the output features to be more discriminative,strengthens the useful information,and suppresses the useless information.?2?A non-local description feature based on feature correlation is proposed.The purpose is to extract non-local relevant information of image features,make the image feature description have a complete representation,and obtain the common features in the class for easy classification.Next,the correlation between features will be the study.By convolving the attention-enhancement feature map to output image features of different regions,and then matrix multiplying the image features of different regions,the spatial dependence between any two positions in the image features is obtained,breaking through the convolution the limitations of the local area of operation to obtain non-local information of image features,referred to as non-local description features.Finally,the soft distribution weights output by Attention-NetVLAD and the non-local description features are subjected to element point multiplication,and the feature correlation between each cluster center area is output,which is serially spliced into the local aggregation feature vector output by Attention-NetVLAD,which is optimized Feature description,find common areas in the image class for easy classificationFinally,the proposed VLAD algorithm combining attention and non-local description features was tested experimentally on the Image Netdataset and the UCF101 dataset.Compared with the existing NetVLAD^[17],Ghost VLAD^[32],and Action VLAD^[33]algorithms,all It has obvious advantages in recognition accuracy.