Multi-feature Fusion Algorithm For Fine-Grained Image Retrieval

The difficulty in fine-grained image retrieval tasks is that objects belonging to the same sub-species usually differ in posture,scale,and background,and the objects of different sub-species may be very similar in these respects.Compared to general image retrieval tasks,fine-grained retrieval meets even greater challenges.At present,there are few studies on fine-grained image retrieval,so effective techniques are urgently needed.This paper aims at improving image feature representation,proposes multi-feature fusion algorithm for fine-grained image retrieval.The retrieval framework is divided into three phases: the coarse retrieval phase is used to narrow the search space and locate the object;the fine-grained retrieval phase uses multi-feature fusion ideas to perform fine-grained image retrieval;the query expansion phase is used to further improve the retrieval accuracy.(1)This paper designs a new constraint to select discriminative patches,which makes use of the irregular but more accurate object region to select the discriminate patches.(2)At this stage,a large number of patch features are mostly clustered or summed and averaged.But due to the uncertainty of the clusters number and the different importance of each patch,there is still have room for improvement.In this paper,after extracting patch features from the convolution layer,based on the fact that the activation value of object is larger,a weighted max-pooling aggregation for patch features is proposed to weaken the possible residual background information and retain as much effective object information as possible.(3)In the current methods of using multi-level features,a simple concatenated method is generally used,which lack of deep mining of intrinsic correlation between features.Therefore,in this paper,deep belief network is introduced to effectively fuse the multi-level features,which capture the intrinsic correlation and rich complementary information contained in the multi-level features.This tragedy can obtain a better image representation.Finally,the effectiveness of proposed fine-grained image retrieval algorithm is proved through compared with other fine-grained image retrieval methods.The experimental results show that the proposed fine-grained image retrieval framework can comprehensively represent fine-grained image features and achieve better retrieval performance.