【摘要】 为解决相互学习中两个小型学生网络在训练过程出现图像局部信息丢失，造成单个学生网络接收精炼特征图不完整的问题，提出基于知识蒸馏的特征精炼相互学习方法。通过对每个学生网络配备辅助网络，提供精炼特征映射和软标签，保留特征的局部信息，为学生网络之间相互学习传递精炼提取知识，提高视觉任务适用性。实验结果表明，在细粒度视觉分类任务公开数据集和基准数据集下的准确率与深度相互学习和自我知识蒸馏方法相比均有明显提升。更多还原

【Abstract】 To deal with the problem that two small student networks lose local information of image during the training process,which results in a single student network receiving incomplete refined feature map,a feature refining mutual learning method based on knowledge distillation was proposed.Each student network was equipped with an auxiliary network to provide refined feature mapping and soft label,retain the local information of features,transfer refined extracted knowledge for mutual learning between student networks,and improve the applicability of visual tasks.Experimental results show that,compared with the deep mutual learning and self-knowledge distillation methods,the accuracy under public datasets for fine-grained visual categorization task and benchmark datasets is significantly improved.更多还原