Weighted Multi-feature Fusion Algorithm for Fine-Grained Image Retrieval

The purpose of Fine-Grained Image Retrieval is to search images that belong to the same subcategory as the query image. In this paper, we propose a Weighted Multi-Feature Fusion Algorithm (WMFFA) to improve image feature representation for fine-grained image retrieval. Firstly, we designed a new constraint to select discriminative patches, which makes use of the irregular but more accurate object region to select the discriminate patches. Secondly, based on the fact that the activation value of an object is larger in the convolution layer, a weighted max-pooling aggregation method for patch features is proposed to weaken the possible residual background information and retain effective object information as much as possible. Current methods of using multi-level features generally use a simple concatenated method, which lacks deep excavating intrinsic correlation between features. Therefore, thirdly, we introduce the Deep Belief Network to effectively fuse multi-level features, which captures the intrinsic correlation and rich complementary information in multi-level features. Experiments show our WMFFA framework achieves significantly better accuracy than existing fine-grained retrieval and general image retrieval methods.