Multi-object Recognition And Pose Detection Of Micro-assembly System Based On Convolutional Neural Network

With the developing of modern science and technology,human research has gone deep into microcosm,meanwhile,miniaturization is going to be a trend in many areas.The scale of working space under microscopic world is extreme small,and the size of operation object is tiny.As a result,it is beyond the precision limitation using conventional scales method.Therefore,micromanipulator,as a kind of approach to research the micro world,has been paid more attention by researchers all over the world.The ICF fusion targets automated assembly experiment requires micro vision which acquires the information about the location and posture of the objects and feeds it back.In this paper,we study the problem of multi-object recognition,location and pose detection based on micro-vision.The precision of object location and pose detection produces an effect on subsequent assembly operation.However the traditional object detection method is poor effect to partially occluded object identification and is unable to adapt to complex scenarios.Recent years,convolutional neural network models trained by deep learning algorithm have achieved great success on object identification and detection.Aiming at the characteristics of micro assembly system target this paper trains the network based on deep learning and using the Faster R-CNN object detection architecture and ZFNet convolutional neural network.It is difficult to obtain micrograph samples of micro-operating system,therefore the available data is far from enough being directly put into the convolution neural network for training.This paper use virtual sample generation method to produce a sufficient number of samples from collected samples.The diversity of samples can meet the demand of actual micromanipulation system task.For example,some of the generated pictures will overlap with each other,have a random attitude angle and simulate color changes in different light intensity.Transfer learning is an important approach to enhance the convolutional neural network generalization.We initialize the micromanipulation system target detection network with a pre-trained network and the network is further trained with the target data of micro-operating system.We analyzed using convolutional neural network principle and method for pose detection,and a pose detection sub-network is added to the Faster R-CNN framework.The post detection network after being trained is tested on micro-images in Micro-operating system.The experimental results show that using convolutional neural network trained by deep learning method can effectively identify and detect the partially occluded objects.Compared with the traditional method,this method has strong adaptability to environment and speediness,also practical application value.