Design And Implementation Of Object Recognition And Positioning System Based On Deep Learning

Industrial robots are a key component of the intelligent manufacturing industry.With the improvement of automation level,the research on robot technology has also received more and more attention.And machine vision technology is undoubtedly one of the most popular research fields.Guiding industrial robots through visual technology is of great significance to increase the level of intelligence in the production process,reduce manual intervention,and promote the rapid development of intelligent manufacturing.Recognition and positioning of objects in actual application scenarios is one of the most common visual tasks.In this paper,we design and implement an object recognition and positioning system based on deep learning for the tasks of automatic recognition and target positioning of objects in a specific production process.In view of the many nonlinear distortions in the camera calibration process that need to be expressed by complex mathematical models and the task of mapping twodimensional pixel coordinates to three-dimensional coordinates in the visual system,we propose an end-to-end BP neural network to fit complex non-linear relationship,and without the help of homography matrix could achieve the direct mapping of coordinates.The experimental results show that the error could be controlled within 0.5mm,which could fully meet the actual application needs.Aiming at the problem that the edges generated by the RCF edge detection model based on the convolutional neural network are too rough,we propose a method of model optimization.In order to solve the problem of improving the resolution of the feature map,we used sub-pixel convolution that could generate finer edges and have lower time complexity instead of deconvolution.According to the intuitive experience of the edge map and the model evaluation results on the standard data set,it could be proved that this optimization strategy could generate a more accurate and clear edge prediction map.Aiming at the task of object recognition and positioning,we propose a template matching method based on edge features.Combining the ORB algorithm and FLANN algorithm could obtain the centroid of the object to be tested.For the integration task of this vision system,we use the industrial control software Kingview to achieve unified control.The feasibility of the vision system has been proved through experiments and practical applications,and the research content of this paper has certain theoretical significance and application value.