Three-dimension Workpiece Recognition Based On Deep Learning

With the development of industrial automation and intelligence,industrial robots are widely used in most aspects of industrial production and processing.In the process of workpiece processing and sorting,the types of workpieces need to be identified first.At present,most of the industrial robots use traditional visual algorithms such as template matching to complete the workpiece recognition task.As the structure of the workpiece becomes more and more complex,the traditional recognition method has been difficult to complete the recognition task.In this thesis,the deep learning algorithm based on RGB-D image and the deep learning algorithm based on multiple views are respectively adopted to complete the recognition task of complex 3D workpiece for the problems of poor universality and poor robustness of the change of workpiece pose,environmental noise and other factors.The research on 3D workpiece recognition algorithm is mainly carried out with 84 kinds of polished workpieces.The main contents are as follows:(1)Firstly,we build the dataset needed in this topic.The Phong illumination model is used to generate a 2D multi-view rendering dataset based on the workpiece CAD model,and the RGB-D dataset of the actual workpiece is constructed by the Kinect v2 camera.Aiming at the problem of background noise in the actual workpiece dataset,the workpiece object is segmented by the Mask R-CNN network to achieve the purpose of isolating the background noise,so as to improve the quality of the actual workpiece dataset.(2)In the 3D workpiece recognition task,this thesis firstly uses the multi-modal convolutional neural network based on RGB-D image to identify the task.By registering the color image and the depth image and adjusting the image size,the workpiece target rectangular region is segmented from the original workpiece dataset and processed into a fixed scale.Depth image cavities and noise problems are solved by pixel filter and median filter.Then the depth image is encoded into a color image as a network input data by the color map mapping method,and the VGG-M network pre-training model was used for network training.The experimental results show that the recognition accuracy can reach to 93.2%when the viewpoint and pose are relatively fixed.(3)Because the multi-modal neural network is less robust to viewpoint and pose,this thesis uses a multi-view convolutional neural network algorithm and improves it.The multi-view dataset based on CAD model and the actual workpiece dataset were respectively used to train the MVCNN network Experiments show that the MVCNN network has poor recognition accuracy on the actual workpiece dataset.In this thesis,a three-stage recognition algorithm is proposed for this problem,which further restrains the noise by instance segmentation and image preprocessing of the workpiece target.Finally,the MVCNN network is used for recognition,and the recognition accuracy reaches 90.8%.For the problems of feature repeated extraction and mask mapping error in the three-stage recognition algorithm,the Mask-MVCNN combined network is proposed.By directly applying the mask to the output feature map of the Mask R-CNN network,the feature map of the workpiece target is obtained.And as the input to the classification phase of MVCNN network.Finally,the effectiveness of the algorithm is proved by experiments,and the recognition accuracy of the algorithm reaches 92.4%.(4)Finally,we compare the two methods in this thesis.The multi-modal neural network based on RGB-D image has a recognition accuracy of 93.2% in certain cases.At the same time,only one image acquisition is needed for the recognition target,the recognition speed is fast,but the robustness to the viewpoint change and pose is poor.The recognition accuracy of Mask-MVCNN combined network is slightly less than the method based on RGB-D image.But it is robust to the above factors,and is more suitable for recognition tasks with complex3 D workpiece.However,due to the need to collect and extract features from multi-view images,the recognition process is more time consuming.