Grasping Point Recognition Of Irregular 3D Objects Based On Improved Mask RCNN

In recent years,due to the increase in labor costs,every industry is seeking ways to upgrade the system.The use of robots for automatic grabbing on automated production lines can effectively reduce production costs and increase production efficiency.In the process of upgrading the automated production line of factory,it is especially important to use image technology to automatically identify targets and grasp it.Convolutional neural networks and machine learning have made important improvement in image processing and image recognition.Compared with traditional visual algorithms,convolutional neural networks is characterized by high generalization and high recognition accuracy.Based on the demand of automation industry upgrade and the development of machine learning theory,a network model that can identify the grasping point of objects is designed to solve the problems,which the robots grasp the target with low accuracy.This paper proposes a two-level network to generate grasping points and also completed network's training and testing.The network structure and training detection method of convolutional neural networks are studied before designing the network.Using the grad-CAM method to design experiments,it is verified that the weights in the general grasping point generation network are more affected by the texture information.According to the characteristics of texture features in shallow network of convolutional neural networks,it is proposed that increasing edge information can increase the success rate of grasp.Consider changing the image channel information will cause feature loss and reduce the probability of grab accuracy.The design experiment uses random image channels to replace one channel or two channels in the RGB image as training datasets,respectively training the network and verifying the accuracy of the network.It is concluded that the loss of channel information leads to a decrease in accuracy,but when the mask image and the depth image are replaced in the channel of the training image,the accuracy is reduced by only 9.3% with respect to the complete image information.The feature information can complete the classification task.Through the above experiments,a crawling point detection network with two-level network was designed.The first-level network uses ResNet101 as the feature extraction network,the RPN as the target classification network,and the deconvolution layer as the mask to generate the Mask RCNN network.The primary network generates classification information and maskimages of the target.After obtaining the mask image and the target classification,the gray image,the mask map and the depth map are combined to form a new three-channel image,which is sent to the second-level grab frame prediction network.The image classification result predicted by the first-level network is spliced as a unique heat code to the feature generated by the convolution layer.The RPN network is used as the grab box regression prediction,and the fully connected layer is used as the classification output of the grab angle.The network eventually got 87.5% of the grab point prediction accuracy.Compared to using RGB three-channel images as a training set,the accuracy of the network is increased by 11.3%.Design a network generalization experiment with eight additional goals to verify that the guessing of the crawling point can also be generated by the network when detecting targets other than the identified categories.Finally,build the simulation environment to simulate the robotic grasping action,and the conclusion that the secondary network can accurately capture the target in the simulation environment is obtained.