计算机与数字工程2026，Vol.54Issue(3)：595-600,622,7.DOI:10.3969/j.issn.1672-9722.2026.03.002
基于注意力机制的多视觉任务混合学习策略
A Mixed Learning Strategy for Multi-visual Tasks Based on Attention Mechanism
摘要
Abstract
With the increase of computer vision task complexity and data scale,single-task vision models tend to ignore the feature information of other tasks,resulting in redundant training and poor model generalization ability.In response to these prob-lems,this method proposes a mixed learning strategy for multi-visual tasks based on attention mechanism,so that the model en-ables to extract the key features of each task while sharing features in multiple tasks.At the same time,this method proposes an adaptive weighting scheme to prevent individual tasks from dominating the multi-task learning,so that all the tasks involved in the learning can be better converged.Experiments on the public data sets such as PASCAL VOC 07+12 and miniImageNet show that compared with the single-task baseline model,the prediction accuracy of this strategy is improved by 1.7%and 2.9%in image clas-sification and image detection,respectively,and the amount of model parameters is reduced by 48.2%.This method solves the prob-lems of parameter redundancy and convergence imbalance in the application of conventional multi-task learning in the field of com-puter vision.关键词
多任务学习/注意力机制/卷积神经网络/自适应任务权重Key words
multi-task learning/attention mechanism/convolution neural network/adaptive weight分类
信息技术与安全科学引用本文复制引用
王康,施必成,顾越,阮俊豪..基于注意力机制的多视觉任务混合学习策略[J].计算机与数字工程,2026,54(3):595-600,622,7.基金项目
国家重点研发计划项目(编号:2017YFB1400704)资助. （编号:2017YFB1400704）
