Instance Retrieval Based On Deep Convolutional Neural Network

In recent years,with the rapid development of social media,people have produced a large number of images and videos in their daily lives.These media materials need to be processed by machines quickly and accurately to extract key information and features.The instance retrieval technique has very practical value in this respect.Especially with the popularization and development of the technology of deep learning,more and more information extraction methods are available for images.Therefore,this paper explores the instance retrieval technology based on deep convolutional neural network.The main contents are as follows:A novel image representation algorithm based on Gram matrix is proposed.By introducing Gram matrix,the second-order aggregation method of convolution neural network feature map is designed.The response relationship between different channels of the feature map is considered,and a more effective global feature representation is obtained.Optimize location of the target candidate in the image.With the continuous iteration of key point matching and affine transformation,the position of the target box in the training dataset is automatically obtained.Based on this dataset,more accurate target regions can be obtained by training the region proposal network.More stable regional features can be extracted,so that more effective global features can be aggregated.Introduce feature learning to learn global feature representations in depth network optimization.Based on the triplet loss function,three-stream Siamese neural network is designed and the network parameters are further optimized through training.The R-MAC and PCA operations are integrated into the network layers,making the network can be trained by end-to-end manner.This also helps the test phase.The image only needs one time network feed-forward operation and the output is a valid global feature representation.Experiment results on publicly available image datasets such as Oxford,Paris,Holidays,Flickr and others show that the above algorithm effectively improves the accuracy of the instance retrieval.