Intelligent Workpiece Recognition,Detection And Grabbing Based On Deep Learning

Under the background of the continuous improvement of industrial automation,the grasping system of robotic arm based on visual system is a key link to realize industrial automation production.Aiming at the problems such as inadequate accuracy of workpiece recognition and detection,and difficulty of traditional feature extraction and multi-workpiece location in the process of robotic arm automatic grasping,this paper addresses automatic workpiece recognition and detection model based on deep learning especially on Convolutional Neural Network(CNN)for the robotic arm grasping system,thus resulting in an intelligent grasping system of manipulator.The main contents of this paper are as follows:(1)For smart manufacturing,a workpiece recognition model based on fog and cloud computing is proposed,and Workpiece Recognition-Network(WR-Net)is formed based on improved classification model Alexnet.The experimental results show that WR-Net avoids network delay and guarantees real-time operation of the model by cloud training and fog deployment.After trained in cloud,the WR-Net is downloaded in fog,and its recognition accuracy reaches 99% for 100 different workpieces.The accuracy of WR-Net is improved by 1% compared with original Alexnet,its model parameters are reduced by 25%,and its real-time performance is guaranteed by avoiding network delay.(2)For the problem that the detection network has low detection accuracy for small-sized targets,on the basis of a single Shot Multibox Detector(SSD),a new feature fusion structure is added to further fuse the low-order and high-order features of the image,thus an improved single Shot MultiBox Detector for Workpiece(ISSD)is proposed.The experimental results show that the detection accuracy of ISSD for five different types of workpieces is 99.2%,which is 2.9% higher than before.ISSD takes 0.026 seconds for the detection of a single picture.(3)In view of the difficulties in deploying the CNN model with relatively limited hardware resources and the high real-time requirement of the recognition and detection model in industrial systems,this study compresses the CNN by means of network pruning,weight quantization and separation convolution,and tests the proposed WR-Net and ISSD models respectively.Experiments show that the WR-Net can achieve lossless compression by subtracting 50% of the parameters via network pruning,and lossless compression can be achieved by reducing the model size by 75% after quatized,but the model recognition speed is not improved.ISSD can reduce its size to 26% of its original one by deconvolution,and improve the detection speed by 42% with only accuracy loss of about 8%.(4)A vision-based intelligent grasping system for robotic arms is constructed by an industrial camera,a notebook computer,and robotic arms to verify the ISSD workpiece detection model.The manipulator grasping system obtains the workpiece image through the calibrated industrial camera,and then the trained ISSD model identifies and detects the workpiece.Finally,the computer controls the manipulator to successfully grasp multiple target workpieces after terminal coordinate conversion.