Multi-object Attitude And Recognition For Automatic Production Line Based On Vision

Automatic production line,the target recognition is an important topic of parts,in this study for automatic production line for special electronic components of the component inserter multi-objective parts recognition attitude judgment puts forward a multi-objective and gesture recognition strategy,aims to solve the component inserter before picking up the target element,the material plate scattered parts recognition and pose estimation problem.In this study,2d image of the target component is collected,the image data set is expanded,the target recognition network is built and the attitude estimation of the component is studied,which provides the theoretical and technical basis of vision for the research of the special-shaped electronic component machine.Firstly,this paper proposes a data set augmentation strategy based on generative adversarial network,and designs a component 2d image generator based on real data set through deep convolution generative adversarial network.The influence of parameter adjustment on convergence in deep convolutional generative adversarial networks is emphasized.In order to evaluate the quality of images generated by deep convolutional generative adversarial network,this study introduced the maximum mean difference evaluation index to select high-quality images and obtain inference data set.The inference data set and real data set were combined to generate a true-inference data set with a large sample size.Secondly,based on YOLOv4 deep learning network,a recognition network model is derived based on the true-inference data set as training set,and the influence of network parameters on small components and network convergence is discussed.In order to evaluate the performance of the model,the average accuracy mean is introduced.The effectiveness of the multi-target recognition strategy in this study is proved by comparing the average accuracy of the training model with the real inference data set and the single real data set.Finally,the attitude recognition of the target component is classified as a kind of image registration problem,and the optimal particle swarm optimization algorithm is used to calculate the attitude of the target component.This research from two aspects of speed up the convergence rate of the particle swarm optimization algorithm,the first is through the Canny edge detection and minimum Oriented bounding box to a rough estimate of target components gesture to reduce the search range of particle,the second for the particle swarm optimization algorithm itself was improved,the inertia weight index of the dynamic.An effective attitude recognition strategy of target element is obtained,and the strategy steps of multi-target attitude recognition algorithm adopted in this paper are summarized.