Dynamic self-adaptive binary hierarchical vocabulary tree image retrieval method

Abstract

The invention discloses a dynamic self-adaptive binary hierarchical vocabulary tree image retrieval method, comprising the following steps: (1) extracting features; (2) constructing a binary hierarchical balance vocabulary tree; (3) selecting two centroids in the space as segmentation points each time, and segmenting the whole space by a hyperplane perpendicular to a straight line passing throughthe two centroids until all the points are segmented; (4) describing each subspace through the centroid of the subspace; dispersing the extracted N descriptors on a node of a binary tree of which thedepth is log2N + 1; (5) performing feature extraction on a picture to be retrieved to obtain a descriptor set P * M which represents P feature vectors; (6) sequentially traversing the binary hierarchybalance vocabulary tree by the P feature vectors and retrieving; and (7) according to the inverted index, each node having a picture ID corresponding to the node, and taking intersection of picture ID sets corresponding to leaf nodes obtained by P feature vectors through retrieval traversal to obtain a picture ID to be retrieved.

Description

technical field [0001] The invention relates to the field of image retrieval, especially image retrieval based on content. Specifically, it is a data structure of a dynamic self-adaptive binary hierarchical vocabulary tree. Background technique [0002] In recent years, Convolutional Neural Networks (CNN) have achieved great success in pattern recognition and computer vision, in areas such as image classification, object detection, instance segmentation, etc. Despite CNN's success, it still has some serious problems. An example is the presence of adversarial examples, when we add small noise or make small changes to the initial samples, the CNN will predict differently for these samples with high confidence, although visually it is difficult for us to find any meaningful to change the picture. Another example is CNN's ability to reject, when a sample from an unknown class is fed to the CNN, it will still assign the sample to a known class with high confidence (should be c...

AI Technical Summary

Problems solved by technology

This explains why the rejection ability of CNN is very poor; secondly, from the perspective of representation learning, the learned representation of CNN is linearly separable, such as figure 1 As shown, under this representation, the distance between classes is sometimes even smaller than the distance within a class, which greatly reduces the robustness of CNN in real complex environments

Images