Method and device for building processing model applied to natural image

Abstract

The invention provides a method and device for building a processing model applied to a natural image, and relates to the technical field of deep learning. The method comprises the following steps: collecting the natural image, obtaining a sample image, and taking the sample image as a data set; segmenting the data set to obtain a training set and a test set, wherein the intersection of the training set and the test set is an empty set; supplementing three-dimensional information for a training set, mapping the two-dimensional image data into three-dimensional point cloud data, taking the three-dimensional point cloud data as input data for training, training by adopting a deep learning mode, and obtaining a processing model for generating a classification result of a natural image to be recognized. The feature embedding space is expanded by restoring the three-dimensional prototype of the two-dimensional image, global and local features are combined at the same time to carry out rapid few-sample learning and prediction, rapid learning and prediction can be carried out under the condition that only a small amount of data exist, and the invention can be better applied to the aspect of natural image classification by combining the characteristics of various features.

Description

technical field [0001] The invention relates to the field of deep learning technology, in particular to a method and device for building a processing model applied to natural images. Background technique [0002] Deep Neural Network (DNN) is a technology in the field of Machine Learning (ML). At present, deep neural networks have achieved breakthroughs in the field of computer vision through abundant and accessible labeled data, but with the increase of labeling costs, the development of deep learning has gradually slowed down; It is very good to summarize the data, but this is obviously contradictory to the current development of deep learning. For these reasons, there is currently a greater focus on learning less in deep neural network techniques. Among them, meta-learning (Meta Learning, ML) refers to the learning of the learning mechanism of the individual, which involves the question of how the individual acquires the function that it relies on for learning. The purpo...

AI Technical Summary

Problems solved by technology

[0005] The present invention provides a method and device for building a processing model applied to natural images, which is used to solve the defect in the prior art that the deep neural network only learns the relationship between instance-level or local information, and cannot predict activities more accurately. Realize fast few-sample learning and prediction by combining global and local features at the same time, and can learn and predict quickly with only a small amount of data

Images