An image retrieval algorithm based on multiple convolutional features of RPN and weighted cosine similarity
Xinhua Liu, HU Gaoqiang, Xiaolin Ma, Hailan Kuang · 2018
Aiming at solving the difficulty of feature selecting and the poor retrieval result in the image retrieval, we proposed an image retrieval algorithm which was based on multiple convolutional features of RPN and weighted cosine similarity. We used the deep learning network RPN to extract the multiple convolutional features; and ranked the images by the weighted cosine similarity we proposed. To validate this algorithm, we used Oxford Buildings 5k and Paris Buildings 6k as two image retrieval datasets. As a result, our algorithm improved the mAP by 3% ~ 6% and had better retrieval and ranking results.