Multi-view target recognition and retrieval method and device based on incremental learning

Abstract

The invention discloses a multi-view target recognition and retrieval method and device based on incremental learning, and the method comprises the steps: dividing a view sequence of a three-dimensional model into a plurality of task sequences by taking a single category or several categories as a unit, and inputting the task sequences into a neural network in a data flow manner; adding a stability module based on feature distillation to the neural network for restraining evolution of old category target features; adding a plasticity module based on knowledge distillation to the classifier for improving the adaptability to a new category target; and extracting each view feature in the view sequence by using a neural network and generating a view feature sequence, fusing the view feature sequences into a feature descriptor, and performing multi-view target identification and retrieval by using the feature descriptor. The device comprises a processor and a memory. According to the method, the incremental multi-view target recognition and retrieval precision is improved by utilizing the stability module and the plasticity module.

Description

Technical field [0001] The present invention relates to view sequences, multi-view target identification and retrieval, and inventory of incremental learning, in particular, to a multi-view target identification and retrieval method, and apparatus for incremental learning. Background technique [0002] Three-dimensional model classification and retrieval is one of the basic technologies in computer vision and multimedia field, which can be directly applied to automatic driving, industrial manufacturing and digital entertainment. [1] . In recent years, the number of three-dimensional objects has grown rapidly, making the multi-view target identification and search methods have been greatly concerned. At the same time, a large number of work is committed to constructing a different descriptor [2] . Existing method [3] Multiple views are typically obtained by placing a virtual camera around the three-dimensional object, and then the features of each view are extracted by a neural ne...

AI Technical Summary

Problems solved by technology

[0009] The present invention provides a multi-view target recognition and retrieval method and device based on incremental learning. Inspired by the successful application of attention-based distillation in the field of two-dimensional image classification, the present invention adds a convolution layer in the middle of the backbone network Based on the stability module of spatial pooling feature distillation, a knowledge distillation plasticity module is added to the output of the classifier, which successfully solves the problem that existing methods cannot adapt to new target categories online or have simple and rough structures when processing multi-view target data streams. and other defects, and on this basis, increased attention to the "stability-plasticity" balance of the incremental learning network, using the stability module and plasticity module to improve the accuracy of incremental multi-view target recognition and retrieval, see below for details describe:

Images