Fine-grained image retrieval method based on self-attention mechanism weighting

Abstract

The invention relates to the technical field of image retrieval and computer vision, in particular to a fine-grained image retrieval method based on visual attention mechanism weighting. The method comprises the following steps of: image preprocessing: setting the length of the longest side of an image to be 500 pixels; feature extraction: inputting the image into a convolutional neural network, and then selecting and outputting the features of the last convolutional layer; target feature selection: firstly, optimizing a local activation graph, and then selecting a local feature vector according to an activation graph result, so as to realize more accurate target feature selection; feature weighted aggregation: evaluating the importance degree of each feature, so as to enable the weightedfine-grained local features to still be embodied during pooling aggregation and improve the precision of fine-grained retrieval; and performing image retrieval, and calculating cosine similarity between the characteristic vectors of the queried image and a database image. An image feature extraction and coding detail graph is shown in figure 1. According to the method, fine-grained image retrievalcan be realized, and the retrieval accuracy is improved.

Description

technical field [0001] The invention relates to the technical fields of image retrieval and computer vision. In particular, it relates to a fine-grained image retrieval method based on visual attention mechanism weighting. Background technique [0002] Although image retrieval has achieved remarkable performance, it is still a challenging problem at the fine-grained image level. Compared to general image retrieval tasks, fine-grained image retrieval methods should be able to localize and express subtle visual differences within subcategories. For example, given a query image that contains a subcategory of a base category, such as the "black-footed albatross" subcategory of birds, we should return images from the database that are in the same subcategory as the query, rather than simply returning an arbitrary bird class image. [0003] Existing deep learning-based image retrieval methods can be divided into two groups according to whether they need to train models on new d...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the shortcomings of existing methods such as inaccurate target feature selection and high time-consuming calculation of visual attention methods, the present invention proposes an effective method to create a powerful feature representation for fine-grained image retrieval

Images