Image Retrieval And Zero-shot Object Detection Based On Deep Metric Learning

Deep metric learning is based on feature extraction and similarity learning,which purpose is to achieve a high-dimensional embedding space where examples from the same class are closer than examples from different classes.Compared with traditional image classification tasks,deep metric learning has higher requirements for feature generalization.For categories that have not been seen during model training,it must also have the ability to distinguish them in the feature space.Metric learning based zero-shot object recognition is achieved by predicting the visual feature information of unseen categories and combining a large number of categories of semantic information,which can be easily extended to zero-shot object detection.Image retrieval based on deep metric learning has attracted wide attention due to its practical application value in many fields such as face recognition,face verification,product recognition,and person re-identification.In this work,we mainly explore how to improve the performance of image retrieval in the two stages of model training and feature post-processing.In the model building and training phase,we focus on designing the network structure and loss function.We designed a network based on batch normalization module and a variety of attention modules such as spatial attention module,channel attention module,so as to verify the role of different attention mechanisms in metric learning.We designed a loss function that takes into account hard example mining and the overall data distribution,therefore validating the role of multiple hard example mining algorithms in metric learning.In the feature post-processing phase,we perform feature transformation on the feature vector in the feature space.We proposed the L2C feature transformation method to further mine discriminative information in the feature vector,which achieved a significant performance improvement on the coarse-grained dataset.Our proposed QE_L feature re-ordering algorithm can effectively use adjacent approximate features to improve retrieval performance.Feature post-processing can reduce the impact of noisy data and make full use of the features of the neighbor retrieval results to further improve the performance of image retrieval.In order to visualize the similarity relationship between different image pairs,we propose a cascaded class activation map and also a self-supervised object feature keypoint detection and matching method based on the attention mechanism.At the same time,we use a simple but effective similarity activation map to directly visualize the similarity response between two image pairs.Finally,combining the semantic information and the visual feature information of images,the visual feature estimation of unseen categories is realized based on the classic two-stage object detection framework,thereby further implementing zero-shot object detection.Experimental results show that the performance of the proposed algorithm has been greatly improved on multiple datasets.