Robot data collection iterative training method and system based on active learning technology, and storage medium

Abstract

The invention discloses a robot data collection iterative training method and system based on an active learning technology and a storage medium. The method comprises the following steps: S1, taking labeled picture data in a preset proportion as training data, and taking picture data in a remaining proportion as test data; s2, establishing a supervised deep learning model, and training the deep learning model by using the training data to optimize the supervised deep learning model; s3, detecting the confidence coefficient of the detection result of the supervised deep learning model by using the test data; s4, formulating a strategy for collecting a rough service data set by the robot, and collecting the rough service data set; and S5, introducing an active learning course to guide active learning process mining to carry out manual annotation on low-confidence samples in the collected rough service data, and by means of the active learning technology, a semi-supervised robot data collection iteration system is realized, the effectiveness of data collection is greatly improved, and the cost of manual annotation is reduced.

Description

technical field [0001] The present invention relates to the technical field of image labeling, and in particular to a robot data collection iterative training method, system and storage medium based on active learning technology. Background technique [0002] With the development of deep learning, the industry is increasingly relying on supervised learning technologies such as target detection and semantic segmentation. Such technologies need to be supported by specific labeled data. Generally speaking, the richer the amount of business data, the higher the recognition rate of robot target detection. High, while labeling data needs to rely on manual labeling with certain training experience, but in actual business scenarios, the cost of labeling large batches of data is relatively high. [0003] Active learning is a method to reduce the cost of labeling and improve the quality of data sets through technical means or mathematical methods. The unlabeled data collected and play...

AI Technical Summary

Problems solved by technology

[0002] With the development of deep learning, the industry is increasingly relying on supervised learning technologies such as target detection and semantic segmentation. Such technologies need to be supported by specific labeled data. Generally speaking, the more abundant the business data, the higher the recognition rate of robot target detection. High, while labeling data needs to rely on manual labeling with certain training experience, but in actual business scenarios, the cost of labeling large batches of data is relatively high

Images