Towards The Weakly Supervised Feature Representation Based Instance Retrieval

Instance retrieval refers to the task of retrieving a given specific instance in a largescale database,and providing the location coordinates of the instance.Instance retrieval is widely used in real scenarios such as online shopping and video editing.At present,the main challenge of deep feature based instance retrieval methods is that the training of deep neural networks is too dependent on strongly supervised information,which limits the retrieval task cannot be expanded to large-scale datasets and cannot detect unknown instances.On the other hand,weakly supervised object detection methods train networks only depending on image category labels to realize object localization and feature representations,which is in line with the requirements of instance retrieval to locate instances and extract discriminative features.Therefore,it is a critical research direction to study the weakly supervised feature representation based instance retrieval.According to the above analysis,this thesis proposes the following two instance retrieval algorithms:Firstly,current instance retrieval methods rely on strongly supervised information to locate instances,and the distinctiveness of features is not enough.Meanwhile,inspired from the powerful feature representation ability of attention mechanism,this thesis proposes an instance representation algorithm based on multi-channel attention area expansion for instance retrieval.Specifically,this algorithm is based on weakly supervised information to train networks.In order to make the detector suitable for multi-instance localization,the multi-channel attention area expansion module is introduced.In addition,multi-branch joint training is added to enrich the feature information of middle convolutional layers,which is combined with attention mechanism to enhance the distinctiveness of features.Experiments show that the proposed algorithm has achieved good performance on several instance retrieval benchmarks.Secondly,there are variety of instances in real scenarios.Strongly supervised information based instance retrieval methods are not capable of localizing and representing unknown instances.In this thesis,an instance representation algorithm based on weakly supervised object localization for instance retrieval is proposed.Specifically,the edge information from shallow feature maps is used in this algorithm,and the parameters of some specific network layers are fixed.It combines with a weakly supervised object detection method to localize instances and keep the sensitivities to unknown instances.In addition,in order to enhance the distinctiveness of features,the object-aware weights are introduced to alleviate the problem that directly extracting features from rectangular area may introduce noise.The experimental verification shows that,the proposed algorithm has achieved satisfactory performance on several instance retrieval benchmarks.