基于改进卷积注意力机制的触觉图像识别

Tactile image recognition based on improved convolutional attention mechanism

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%.

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 attention 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 attention 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 classification 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 multimodal sensor dataset,respectively.