摘要
视觉引导的工业机器人自动拾取研究中,关键技术难点之一是机器人抓取目标区域的识别问题,前提是生产流程中合格与瑕疵品的检测识别,是后续分拣抓取的关键步骤。因此论文提出一种结合深度学习和Softmax的区域识别方法,通过分别提取抓取区域的方向梯度直方图HOG特征和中心局部二进制模式特征CS-LBP特征,对融合后的特征采用主成分分析法PCA进行降维处理,以此来训练Softmax分类器进行分类。为此训练了Unet神经网络完成抓取区域的分割操作。然后采用Softmax分类器对Unet识别出的区域进行了二次分类,剔除了干扰区域。最后一步计算掩码,达到对瑕疵区域的精准识别。论文通过对比单一的Unet算法以及多特征融合的Softmax算法,在漏检率、错检率、识别准确率等三个指标上的对比,表明了论文算法的鲁棒性。
One of the key technical difficulties in visually guided industrial robot automatic pickup research is the identification of the robot's gripping target area.The premise is that the detection and identification of qualified and defective products in the production process is a key step in subsequent sorting and grabbing.Therefore,this paper proposes a region recognition method combining deep learning and softmax.By extracting the direction gradient histogram HOG feature of the grab area and the central local binary pattern feature CS-LBP feature,the principal component analysis method PCA is used for the fused feature perform dimensionality reduction to train the softmax classifier for classification.To this end,the Unet neural network is trained to complete the segmentation operation of the grabbing area.Then softmax classifier is used to carry out secondary classification on the area identified by Unet and interference areas are eliminated.The final step is to calculate the mask to achieve accurate identification of the defective area.This paper compares the single unet algorithm and the softmax algorithm with multi-feature fusion on the three indicators of missed detection rate,false detection rate,and recognition accuracy rate,which shows the robustness of the algorithm in this paper.
作者
张新波
朱姿娜
张伟伟
ZHANG Xinbo;ZHU Zina;ZHANG Weiwei(Institute of Vision Measurement,Control and Intelligent Navigation,Shanghai University of Engineering Technology,Shanghai 201620)
出处
《计算机与数字工程》
2022年第5期1142-1146,共5页
Computer & Digital Engineering
