Research On Partial-duplicate Image Retrieval Algorithms Based On The Multi-contextual Clues

The existing content-based partial-duplicate image retrieval algorithms mostly rely on the bag-of-visual-words(BOW)model.Generally,local features are quantized to the visual words to improve the retrieval efficiency.Thus,any two local features from different images are considered as a feature match when two local features are quantized to the same visual word.However,BOW quantization errors result in the low discriminability of local features,this causes many mismatches between images,which will greatly reduce the retrieval accuracy.To solve the above problems,this paper proposes two partial-duplicate image retrieval algorithms based on contextual clues.1)A partial-duplicate image retrieval algorithm based on image representation by multi-contextual clues.To improve the discriminative power of local features,this paper proposes a novel partialduplicate image retrieval approach,which enriches local features' information by spatial and visual contextual clues.In this scheme,firstly each scale invariant feature transformation(SIFT)feature from the images are selected as referential feature and several stable context features for each referential feature are selected in the image.Then it encodes the spatial relationships between referential features and their contexts as spatial descriptors.What's more,it builds region-based color histogram as color descriptors since color is robust to many common image attacks.Finally,it embeds these contextual descriptors that are as supplementary information into the inverted index file to accelerate feature matching.In the image retrieval process,many false matches are detected and removed through multi-contextual clues verification,which can improve the retrieval accuracy.2)A partial-duplicate image retrieval algorithm based on contextual CNN feature post-verification.To improve the retrieval performance,the proposed scheme first obtains all pairs of SIFT feature matches by BOW model,and then takes the advantages of properties of SIFT features(characteristic scale,dominant orientation and coordinates),to detect the potential duplicate regions between images and avoid the inconsistency between detected duplicate regions from partial-duplicate images.Next,contextual convolutional neural network(CNN)features are extracted from these duplicate regions.In addition,to reduce the complexity of features extraction and matching,we also propose a two-stage sum-pooling strategy,which not only generates CNN features with reasonable length,but also preservers the spatial clues of these duplicate regions.Finally,these irrelevant images that are mixed with BOW model are detected and removed by verifying these region-based contextual CNN features.Experimental results show that the proposed approach provides superior accuracy than the state-of-the arts,while achieves comparable efficiency.