Research On Deep Metric Learning Image Retrieval Algorithm Based On Robust Loss And Enhanced Features

With the development of the information age,the image data transmitted by the Internet is increasing by hundreds of millions every day.How to accurately and quickly retrieve the images people need from the huge image library has become a long-term research topic in the field of computer vision.In recent years,deep metric learning methods have achieved good results in the field of image retrieval,but there are still problems such as easy to fall into local optimal solutions and poor anti-interference ability.Based on the framework of the deep metric learning image retrieval algorithm,this paper studies the deep metric learning image retrieval algorithm based on robust loss and enhanced features from the two perspectives of loss function modeling and feature learning.Research shows that the adversarial training method can effectively improve the anti-interference ability of deep neural networks.Based on this,this paper proposes an image retrieval algorithm based on adversarial triple embedding.The algorithm introduces adversarial perturbations to the image features extracted by the deep neural network to generate adversarial triplet samples for the training of the deep neural network,so that the deep neural network can learn appropriate parameters to adapt to the introduced adversarial perturbation,thereby improving the model's performance robustness.The experimental results on six different image retrieval data sets show that compared to other triple loss functions,the image retrieval algorithm based on adversarial triple embedding proposed in this paper can improve the accuracy of image retrieval on each data set.Most of the existing deep neural network models face the problem of poor anti-interference ability,that is,a small disturbance of the input image may cause a large deviation in the output of the model.In order to alleviate this problem,this paper proposes an image retrieval algorithm based on input gradient regularization.The algorithm separately introduces the training loss of the model and the gradient of the image feature relative to the input sample as regular terms to constrain the model,so that the output of the model will not change drastically with the slight disturbance of the input image,thereby improving the robustness of the model.The proposed method has the characteristics of plug and play and can be applied to various types of loss functions.In the experiment,this paper applies the algorithm to four different loss functions and verifies it on four different image retrieval data sets.Experimental results show that on different data sets,the proposed method further improves the accuracy of image retrieval based on different loss functions.The feature expression of images is one of the important factors that affect the performance of image retrieval.In order to improve the expression ability of image features,this paper introduces the convolutional block attention model into the deep metric learning network.Specifically,the convolutional block attention model includes a channel attention module and a region attention module,so that the model pays attention to the features of the region where the target object in the input image is located in the process of extracting features,suppresses the features of the background region,and obtains intra-class semantics correlation,image features that are highly recognizable between classes.The experimental results on three different image retrieval data sets show that the convolutional block attention model can effectively improve the feature expression ability of the image,thereby improving the accuracy of image retrieval.In addition,this paper combines the convolutional block attention model with the image retrieval algorithm based on input gradient regularization proposed above.Experimental results show that the convolutional block attention model is complementary to the algorithm based on input gradient regularization,which can further improve the accuracy of image retrieval.