Research On Deep Enhanced Image Retrieval Technology

With the advent of the era of big data,content-based image retrieval method has shown great practical application value and scientific value.Among many image retrieval algorithms,image retrieval algorithms based on deep convolutional neural networks have received extensive attention from researchers due to their strong nonlinear fitting ability.This paper explores image retrieval models based on deep convolution neural network,which enhances the feature representation ability of convolution layer and full connection layer,as well as the loss function,thus improving the retrieval performance of the model.Firstly,the paper briefly introduces the research background of content-based image retrieval method,and then introduces several representative image retrieval algorithms at home and abroad,and then points out the challenges of current content-based image retrieval algorithms,and finally introduces the content arrangement of this paper.Subsequently,the paper introduces the most representative convolutional neural network structure,then summarizes the deep hash retrieval model,and points out the problems of the current deep hash retrieval model,and then a deep hash retrieval model based on enhanced loss function is proposed,which extends the similarity matrix of images to the range of natural numbers,the ability of approximate hash codes to describe image similarity are enhanced.It not only judges the similarity of images,but also can describe the similarity between images in detail.Moreover,the paper theoretically analyzes the upper bound of the quantization error,and by optimizing the upper bound,obtains a hash code with less quantization error,Finally,the gradient back propagation formula is deduced,which improves the retrieval performance of the model.Then,the third chapter uses the idea of enhanced features to propose a deep hash model that enhances the convolution features.the paper visualizes the information contained in the features of each convolutional layer,constructs the mapping from feature space to pixel space,and the linear combination of the feature maps of the last convolution layer is given,and summarizes the relationship between the distribution of activation values on the last convolution layer feature map and network performance.By enhancing the feature map of the convolutional layer locally,it strengthens the effective features and reduces the redundant features.Then,a back propagation algorithm to enhance the convolution layer is deduced,which improves the network performance.In addition,this chapter designs an effective fold line rectification activation function,and obtains a more robust hash code by optimizing the well-designed loss function,which improves the retrieval precision.Next,the fourth chapter proposes a deep enhanced product quantitative retrieval model.This chapter introduces the shortcomings of the Softmax classifier commonlyused in convolutional neural networks in metric learning,and points out the main factors affecting the performance of the product quantization method.Therefore,this chapter designs a piecewise center loss function,enhancing the features of the fully connected layer,and proposes a convolutional neural network to fit the subspace division and coding process of product quantization,and trains the features with piecewise cluster distribution.The quantization error is effectively reduced and the retrieval performance is improved.Finally,the fifth chapter of the paper summarizes the work of this paper.