首页|基于改进卷积注意力机制的触觉图像识别

基于改进卷积注意力机制的触觉图像识别

扫码查看
为了改善传统轻量化网络对触觉图像全局特征提取能力差的问题,提出一种基于轻量化网络提高触觉图像感知分类的新算法,通过将卷积块注意力模块(CBAM)引入坐标注意力机制(CA)来增强特征信息表达能力.利用CA采取空间全局信息并嵌入通道注意中,使卷积网络能够在较全面的区域捕获注意力权重.结果表明:所提算法优于现有轻量化网络算法;该算法对GelSight数据集、多模态传感器数据集2种触觉图像进行分类识别测试,在分类表现中分辨正确率分别达到了88.2%和94.4%;相比于传统的CBAM注意力模型、自注意力模型(SENet)和仅有LeNet的神经网络,该算法对触觉图像的识别能力在GelSight数据集上分别提高了8.7%、8.7%和3.0%,在多模态传感器数据集上分别提高了13.3%、13.4%和4.8%.
Tactile image recognition based on improved convolutional attention mechanism
To improve the problem of poor global feature extraction of tactile images by traditional lightweight networks,a new algorithm based on lightweight networks is proposed to improve the perceptual classification of tactile images.By introducing the convolutional block attention module(CBAM)into the coordinate atten-tion mechanism(CA),the feature information expression ability is enhanced.The CA is used to take spatial global information and embed it in the channel attention,so that the convolutional network can capture the at-tention weights in a more comprehensive region.The results show that the algorithm outperforms the existing lightweight network algorithms.The algorithm is tested for classification and recognition of two tactile images,GelSight dataset and multimodal sensor dataset,and achieves 88.2%and 94.4%accuracy rates in the classifi-cation performance,respectively.Compared with the traditional CBAM attention model,the self-attentive model(SENet),and the LeNet-only neural networks,the algorithm's recognition ability for tactile images is 8.7%,8.7%and 3.0%higher on the GelSight dataset,and 13.3%,13.4%and 4.8%higher on the multi-modal sensor dataset,respectively.

tactile imagelightweightattention mechanismcoordinate attention(CA)

熊鹏文、陈志远、廖俊杰、宋爱国

展开 >

南昌大学信息工程学院,南昌 330031

南昌大学机器人研究所,南昌 330031

东南大学仪器科学与工程学院,南京 210096

触觉图像 轻量化 注意力机制 坐标注意力

国家自然科学基金资助项目国家自然科学基金资助项目国家自然科学基金资助项目国家自然科学基金资助项目江西省"双千计划"资助项目江西省杰出青年基金资助项目江西省主要学科学术与技术带头人资助项目国家重点研发计划"智能机器人"重点专项资助项目

62163024623731816190317561663027jxsq202320109720232ACB21200220204BCJ230062023YFB4704903

2024

东南大学学报(自然科学版)
东南大学

东南大学学报(自然科学版)

CSTPCD北大核心
影响因子:0.989
ISSN:1001-0505
年,卷(期):2024.54(1)
  • 31