Character interaction-oriented network weight generation few-sample image classification method

Abstract

The invention discloses a character interaction-oriented network weight generation few-sample image classification method. Considering the problem of long tails in a character interaction image data set, a meta-learning framework and an episodic training strategy are used to simulate the situation of less sample image data in a real scene, reduce the image data distribution difference, improve thegeneralization ability, and effectively solve the problem of uneven image distribution in a character interaction task. Introducing semantic information of nouns and verbs of the labels, and generating task-level feature extraction network function parameters to reinforce visual features. TRAINING PROCESS, different small sample tasks are continuously obtained; task data is expressed as a task feature by using semantic and visual fusion information of a category contained in a task, a parameter of a corresponding layer of a target network for the current task is sampled through a parameter generator based on the task feature, and actions and object areas of characters in related images are concerned, so that the image classification capability can be improved.

Description

technical field [0001] The invention relates to an image classification method. In particular, it relates to a few-shot image classification method based on network weight generation for human interaction. Background technique [0002] Deep learning is a hot field in machine learning, with specific tasks such as image classification, object detection, etc. Deep learning technology is based on a large amount of data and large-scale training to simulate or realize human learning behavior to acquire new knowledge or skills. The collection and labeling of these data requires a lot of labor costs. In reality, with the emergence of more application scenarios, we will inevitably face more data shortage problems. However, in the case of relatively few labeled data, neural networks are usually prone to overfitting, which makes the application and effect of deep learning difficult. are restricted. In contrast, humans have the ability to learn from a small amount of data. For examp...

AI Technical Summary

Problems solved by technology

[0009] 1) At present, there is a long-tail problem in the character interaction data set, that is, there are some uncommon combinations, and the categories distributed at the tail have only a small number of samples, resulting in extremely unbalanced samples between categories, which makes the generalization ability of the trained model poor , easy to overfit

How to transfer the knowledge of categories with many head samples to categories with few tail samples poses a challenge to the task of character interaction

[0010] 2) At the present stage, the task structure of meta-learning is relatively simple, and the data sets used for meta-learning training are few, and the evaluation standards are not uniform to some extent, which limits the development of few-sample training based on meta-learning

However, the image size in the meta-learning dataset is small, and the visual scene is single, mainly focusing on things in the middle of the image.

Images