Research On Two-stage Target Positioning Methods Based On Convolutional Neural Network

Target positioning is one of the basic problems and research focuses in the field of machine vision,which has a wide range of applications in workpiece processing,robots and other fields.However,the current visual positioning methods need to be improved in speed,accuracy,robustness and reliability.By improving the target location algorithm,the level of automation and precision manufacturing can be enhanced in the fields like workpiece processing and robots.in this paper,a two-stage convolution neural network is used to optimize and improve the target positioning algorithm,and the experiment is carried out on the precision optical lens auto-swing device based.The contributions of this paper include the following:(1)Based on the review of target positioning,the existing problems of speed,accuracy,robustness and reliability are analyzed.Given the key points of current visual positioning technology,two template matching datasets are constructed to evaluate target positioning algorithms represented by template matching.(2)The influencing factors of the speed and robustness of the template matching method are studied experimentally and improvement methods are present.Firstly,a two-stage variable step search strategy is proposed to speed up template matching.and based on the convolution neural network classification model,an online template matching framework is built to improve the robustness of template matching by adaptively selecting templates,which can overcome the impact of environmental changes such as lighting and background variation.(3)The performance of template matching based on convolutional neural network is experimentally studied.Experiments show that the strong feature extraction ability of convolution neural network improves the accuracy of template matching.and the reliability of template matching algorithm based on convolution neural network is discussed,and a method based on feature point check is proposed to improve the reliability of template matching algorithm based on convolution neural network.(4)A two-stage instance detection framework based on convolution neural network is presented to improve the speed,accuracy,reliability and robustness of workpiece positioning.Through Experiments the two-stage instance detection framework based on convolution neural network shows superior speed and robustness.A new target positioning frame regression method and a region relocation strategy are proposed to improve the positioning accuracy of the algorithm.Then,a combined filtering and validation strategy is proposed to improve the reliability of the algorithm.(5)Based on the machine vision of precision optical lens auto-swing device,by analyzing its visual target positioning requirements during the swing operation,a software and hardware platform for workpiece positioning is established,and the applicability of various target positioning algorithms in this scene is tested through experiments.