Research On Image Retrieval Method Based On Image Principal Part Detection

The last decades have witnessed tremendous advances in Content-based image retrieval(CBIR),there are still some challenges in large-scale image retrieval.First,there is semantic gap between the low-level visual features and high-level semantic concept.Second,the users cannot precisely express the expected visual content for a multi-label query image.Third,retrieval system is difficult to quickly respond to queries in large-scale database.In order to solve the above problems,an image retrieval method based on image principle part detection has been proposed.Object-level feature is extracted by object detection network aimed at eliminating effect of background.Subject detection method is applied to avoid the influence of the non-subject object.Finally,object-level features are mapped to binary hash codes for the improvement of efficiency of retrieval.The main contributions are as follows:(1)An image retrieval method is proposed based on deep object-level features.This method uses the object detection method to extract the deep features of the region proposals.It can effectively avoid the influence of complex background.The retrieved images are ranked by similarity relationship between query and images in database.The similarity is calculated by the category probability and the cosine distance.It can eliminate the impact of object detection errors on the retrieval results and get more accurate results.(2)A subject detection method combined with object spatial relationship and class probability is presented for multi-label images.After object detection,the subject of the image is calculated according to the spatial relationship between the objects and the probability of the classes.The subject is used as the content which users want to retrieval.Thus,the objects which is located at the edge of the image or the smaller area in the query image are ignored.Experimental results show that the subject detection method has a good retrieval effect on the multi-label images and effectively avoid the influence of the non-subject object.(3)A deep feature dimension reduction method is designed by mapping the real-valued features to compact binary codes in a Hamming space.A fully connected layer is added into the object detection neural network to learn the binary hash codes.The activation of that layer is used as hash codes after binarization.Experimental result shows that our deep hash method is better than Instance aware hashing(IAH)method.Compared with the method that use of real-valued vectors,hash method's accuracy decreases slightly while the search time is shorter.