A method and system for image retrieval based on multiple semantic levels

Abstract

The invention discloses a method and system for image retrieval based on multiple semantic levels. The method includes: inputting the image to be retrieved into a pre-training model, and inputting the pixel value of the image and a candidate area with local area information, In the neural network of the pre-training model, the size of the pooling kernel is adaptively adjusted according to the size of the candidate region after mapping the candidate region in the picture to the convolutional feature map output by each convolution layer , to obtain the convolutional feature map of the same dimension; perform region-aware multi-level pooling calculations on the low, middle, and high convolutional layers according to the convolutional feature map, and then obtain feature fusion at different levels through concatenation; according to the feature Combine the results and retrieve the images. The features in the present invention include both local information and global information, as well as visual information and semantic information, thereby improving the accuracy of image retrieval. In addition, a single feedforward calculation operation ensures high efficiency.

Description

technical field [0001] The invention relates to the fields of neural network and image processing, in particular to a method and system for image retrieval based on multiple semantic levels. Background technique [0002] There are two main types of feature extraction methods for image retrieval today: traditional feature-based and convolutional neural network-based. Among them, the traditional feature-based method has weak expressive ability and high feature dimension because the image features are all manually designed. In addition, most of the current convolutional neural network-based methods extract single-layer features in the network, and the features of the fully connected layer are used the most, but these methods ignore the rich local information in the convolutional feature map, and ignore the network. The visual information contained in the middle and low-level feature maps. [0003] Specifically, the existing feature extraction methods for image retrieval are s...

AI Technical Summary

Problems solved by technology

[0004] 1. The disadvantages of traditional features (SIFT) and feature aggregation (VLAD, Fisher Vector) are: high feature dimension, weak feature expression ability, and low retrieval accuracy

[0005] 2. Convolutional neural network (through pre-training and fully connected layer features), the disadvantages are: high feature latitude, feature lack of local information, feature lack of low-level visual information

Images