Research On Instance Search Method Based On Deep Learning

The instance search is proposed by the internationally renowned computer vision competition TRECVID,which means that given a sample to find the picture or video frame containing the sample in the picture data set or a video.The technology is widely used in the field of object and building search and security monitoring.Similarly,local instance search also has certain practicability.In this paper,convolution neural network is used to extract regional features for global instance search and local instance search respectively.In order to solve the problem of low instance search accuracy,a improved Faster R-CNN object detection network is proposed to build a instance search model.In the data preprocessing stage,a variety of data enhancement methods are used to increase the training set,and the balanced sampling strategy is used to increase the training opportunities of small samples.In the retrieval module,the candidate box score and the calculation of cosine distance are combined to improve the accuracy of target location.In the model training module,the initialization method is changed to He initialization method,and the shallow features of the convolutional neural network are fused,and then sent to the later networks to extract high-level features,so as to improve the expression ability of high-level features.In addition,the transfer learning method is introduced to fine-tune the pre-training model on the two open data sets Oxford5k and Paris6k.The experimental results show that the retrieval accuracy rates on the two public data sets reached 0.926 and 0.924,respectively,which were 15.4%and 10%higher than the original instance search accuracy using Faster R-CNN network.Local instance search has important practical significance,this paper proposes to apply the global instance search algorithm to the local instance search task,that is,searching the whole image using incomplete image.Firstly,the local query image library is constructed,and the size of the incomplete query image is further processed by filling.Secondly,the model obtained from the global instance search is applied to the local instance search for retrieval.Finally,the online retrieval function is added,The experimental results show that the accuracy of this method on the two data sets is 0.880 and 0.859 respectively.The accuracy rate is 9.5%higher than the existing methods on the Oxford 5k dataset.The returned result corresponds to the entire image,and can accurately mark the location of the partial query image in the entire building.Finally,the online search function is added,and the search time reaches 5.7s and 7s on the two data sets without coding.The global instance search algorithm proposed in this paper has high retrieval accuracy and significantly improved target positioning.The global instance search algorithm is applied to the local instance search task,the accuracy rate is higher than that of published papers,and it is innovative.