基于注意力机制的全卷积抓取姿态预测方法

A Fully Convolutional Grasp Posture Prediction Method Based on Attention Mechanism

针对机器人抓取姿态预测中准确性和实时性不能平衡的问题,提出一种嵌入空间通道注意力机制(CBAM)的全卷积抓取姿态预测方法。使用深度可分离卷积代替标准卷积以降低模型的参数量;利用残差网络(ResNet)中ResBlock模块的思想,并加入CBAM和Inception模块对ResBlock进行改进,加强模型对特征的多尺度提取,以减少局部信息的丢失;融合浅层特征和深层特征进行反卷积,并引入抓取位置质量作为评价指标,直接预测输入图像中每个像素的抓取姿态,从而提高最终预测的准确性。实验结果表明,提出的抓取姿态预测模型可在满足更高准确率的同时兼顾实时性的要求。

To address the problem of imbalance between accuracy and real-time in robot grasping pose prediction,a fully convolutional grasping pose prediction method with embedded attention mechanism CBAM is proposed.First,the standard convolution is replaced by a depthwise separable convolution to reduce the number of parameters in the model.Second,Inception block and attention mechanism CBAM are added to the ResBlock in ResNet.The improved ResBlock can enhance feature extraction and reduce the loss of local information.Finally,shallow features and deep features are fused.The fused features are deconvoluted to predict the grasp posture for each pixel of the input image.Position quality is added to the final output,which can improve the accuracy of the prediction.The experimental results show that the proposed grasp posture prediction method can meet the requirements of higher accuracy and real-time performance.