The Research Of Aggregation Representation And Image Retrieval Based On Convolutional Neural Network

In the fields of artificial intelligence,computer vision and pattern recognition,image retrieval is a popular technique that it refers to search the need images related to query from a dataset.The national 14 th Five Year Plan proposed one major project that developing a new generation of artificial intelligence to break through cutting-edge basic theories.It includes those innovations in a field of image processing based on deep learning.Thus,image retrieval research is a meaningful foundational task.The critical works of image retrieval are feature extraction and representation.Nowadays,feature representation based on convolutional neural network(CNN)models has become a hot topic.The research community has developed two techniques(e.g.,fine-tuning model and features integration)to provide the high discriminative representation.Features integration techniques are widely studied because of their simplicity and efficiency.Previous studies have shown that representation based on CNN exhibit a phenomenon of “texture-bias”.Furthermore,hierarchical features based on CNN closely related to human visual perception mechanisms.Thus,feature integration can enhance the representation via using various visual features.However,existing methods mainly employ single hierarchical feature or concatenate several representations,ignoring the spatial information.This limits the entire discriminative ability.To address this problem,we proposed four aggregation representation methods based on CNN via simulating the visual mechanisms,to image retrieval.These methods can make a complete use of the discriminative advantages of various features.They can provide the discriminative yet compact representation and effectively improve the retrieval performance.1.Representation based on the contrastive weighting aggregation for image retrieval.A contrastive computation model is proposed to approximately construct the region of the global receptive field,and further to enhance the object region information implied in deep convolution features.After aggregating the enhanced features based on object region,a discriminative yet compact representation is provided to enhance representation by utilizing the spatial layout.The highlight of this method is that we proposed a calculation method of soft weight.The experiments have demonstrated that the proposed method can significantly improve the retrieval performance and outperform the most representative methods.2.Representation based on the object semantic aggregation for image retrieval.This method integrates semantic concepts into the latent object features via constructing semantic attention map based on the spatial-and channel-wise perception,thus,it provides a compact representation and then enhances it by exploiting semantic.The highlight of this method is that we proposed the PCA-p-whitening approach to reduce vector dimensionality.The experiments have demonstrated that the object semantic aggregation representation can well integrate the object and semantic features,and further improve the retrieval performance.3.Representation based on the deep color-orientation aggregation for image retrieval.A computation model based on color orientation is proposed to activate the latent orientation attribute.The representation has an enhanced discriminative power by integrating deep features and color orientation.The highlight of this method is that we explore the color orientation-based computation model to ensure the compatibility between low-and high-layer features.The experiments demonstrated that the proposed representation can aggregate the color features and deep features well,and significantly improve the retrieval performance.4.Representation based on the multi-layer feature aggregation for image retrieval.A computation model was developed for integrating low-layer texton features,middle-layer object features,and high-layer semantic features.After aggregating those features can provide a discriminative yet compact representation to image retrieval.The highlight of this method is that we proposed the difference whitening approach to reduce vector dimensionality,which is easily transferred to new tasks.The experimental results have demonstrated that our proposed representation can integrate the advantages of multi-layer features.Furthermore,it can significantly improve the retrieval performance.The four representations can integrate spatial layout,semantic,color,and texton features,thereby making full use of the hierarchical features in CNN model to describe the image content well.They can combine the advantages of different visual features to enhance representation.On the marine fish image dataset,the four representations have demonstrated their effectiveness.Compared to the baseline and fine-tuning-based methods,they can significantly improve the performance and have the applicability.Furthermore,the multi-layer feature aggregation representation provided the best retrieval performance.