Multi-class multi-scale target sorting method and device based on continuous learning

Abstract

The invention discloses a multi-class multi-scale target sorting method and device based on continuous learning, and the method comprises the steps: achieving the hidden space representation of new and old task features through a generative adversarial network, building a knowledge memory storage warehouse, achieving the quick switching of multi-task sorting by a self-adaptive knowledge migration scheduling module. The training process of a new task is accelerated, the importance degree of model parameters is quantified in combination with a regularization technology, and key parameters are prevented from being covered in the training process. The data replay framework is used for the intelligent part sorting system, the problem of function disastrous forgetting occurring in the continuous learning process of a multi-type multi-scale target sorting model can be solved, the integrity of the function of the sorting model is ensured, continuous learning with dynamically changed targets and scenes is achieved, and the learning efficiency is improved. The system is assisted coordinately and efficiently solve the multi-task sorting problem.

Description

technical field [0001] The invention belongs to the technical field of machine learning, and in particular relates to a multi-category and multi-scale object selection method and device. Background technique [0002] The target picking model is the core technology for the development of intelligent manufacturing and flexible manufacturing. It realizes unmanned production of industrial processing, testing and assembly, and has broad application prospects in the fields of mechanical processing, automobile manufacturing, warehousing logistics and 3C. [0003] At present, when the intelligent visual parts picking system picks multi-category and multi-scale targets, especially when continuously and dynamically training new picking objects, the model parameters are easily covered by the new learning process, and the picking system will have catastrophic forgetting of existing functions. , resulting in a significant reduction in system work efficiency, and damage to the original sy...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the problems in the prior art, the present invention proposes a multi-category and multi-scale object picking method and device based on continuous learning, and adopts advanced continuous learning technology to solve the problem of existing intelligent picking systems dealing with multiple categories and multiple targets. When the goal of the scale is not met, the effective scheduling of the existing picking learning system and the efficient learning of new tasks cannot be adaptively

Images