Centralized Ranking Loss With Weakly Supervised Localization For Fine-Grained Object Retrieval

Fine-grained visual retrieval is different from traditional visual retrieval and requires more detailed feedback.As a new research field,fine-grained visual retrieval faces the following challenges:First,fine-grained visual retrieval has very little difference between different classes.Fine-grained visual retrieval needs to distinguish between the nuances of the target object and the,And the solution to this problem requires a high degree of partitioning metric learning methods and strong representation of the characteristics of description methods.Secondly,due to the difference in shooting time and place,the similar target objects cause large apparent differences between similar individuals,while fine-grained visual search needs to be able to filter such apparent differences and generate approximate feature descriptions to achieve high fine-grainedness.Visual retrieval accuracy.In addition,in the current mainstream metric learning learning method,the calculation of the loss function has a problem of high time and space complexity,resulting in a very time-consuming and labor-intensive training process.Finally,fine-grained visual retrieval methods still have the problem of insufficient sample data with accurate labeling information.Fine-grained visual search requires a large number of samples with accurate positioning information,and the current serious loss of such sample data.In view of the above-mentioned problems in the current fine-grained retrieval method,this paper proposes a unified framework for fine-grained visual retrieval,and based on this framework develops related researches such as depth metric learning and weak-supervised feature representation learning.The main innovations of this article include:1)A loss function based on central point ordering is proposed,which is different from the traditional depth measurement learning loss function.The loss function based on center sorting uses the center point to update the parameters,which not only improves the distinguishability and generalization ability of features,but also reduces the The computational complexity of the algorithm,for the same batch of image input,the training process based on the central sorting loss function is more than 1000 times faster than the training based on other loss functions,greatly improving the training efficiency,making the method Better practicality.2)A weakly supervised feature extraction method is proposed.The method firstly uses the statistical information of the high-level neural network to complete the initial positioning of the target object,and then uses the hybrid Gaussian model and the maximum and minimum graph cutting algorithm to target the initial positioning.The object is refined,and the features are screened and aggregated using refined positioning information to obtain features with high generalization and high discrimination.In this paper,simulation experiments are performed on several related fine-grained visual search data sets.By comparison with other current mainstream methods,we can find that under the same conditions,the proposed method far exceeds the accuracy and training speed in retrieval accuracy.Other current mainstream fine-grained visual search methods.