相似性保持实例检索方法

Aiming at the network training of multi-input samples, a similarity retention instance retrieval method is proposed. Firstly, the input image features are extracted by the convolution structure in the depth network and pooled. Then, according to the benchmark order, the similarity relationship between the low correlation image and the query image is corrected, and the low correlation image contrast loss coefficient is obtained, and the loss value within the loss reference value is retained. The loss value is performed to maintain contrast loss training based on similarity. Finally, the post-training network is used to extract image features for instance-level image retrieval. The experimental results show that the loss comparison function based on similarity is feasible, and the method significantly improves the accuracy of instance-level image retrieval.