Weakly Supervised Class-Agnostic Image Similarity Search Based on Convolutional Neural Network

For vision-based localization methods, the efficient and accurate image retrieval in a database with a large number of images is an important prerequisite and guarantee. The traditional content-based image retrieval method can retrieve database images that have similar details and features with the query image. However, there may be significant differences in semantics between them, which may result in mismatches. The method based on convolutional neural network can avoid such problems to a certain extent. In this paper, we propose an end-to-end weakly supervised class-agnostic image retrieval method based on convolutional neural networks. In the offline stage, the database images are preprocessed to separate the foregrounds and backgrounds, and the foregrounds are clustered for storage. The aim is to reduce the amount of data calculation in the online stage and avoid mismatches caused by backgrounds mixing. In the online stage, the same processing is performed on the query image, and the improved method based on the relational network is used to compare the similarity between the input objects to obtain the optimal retrieval result. The test results on Holidays, Paris 6k, Oxford 5k and three large-scale datasets show that the proposed method has better performance than existing image retrieval methods.