Research On Vision System Algorithm Of Go Robot Based On Machine Vision

Machine vision system is a modern technology which is very novel and serves all fields.It mainly consists of image acquisition and image processing.It mainly studies the analysis and processing of the objects or images collected by the vision system,to meet the requirements of practical application.In the field of machine vision system,people have conducted a lot of research and obtained abundant research results,but in different complex application scenarios,they still face several major challenges,which are the calibration of the vision system,the complex background of the image collection,the different illumination and the target occlusion.How to calibrate the robot vision system and detect the target or category we need in the collected image,and the detection result is not affected by the target position,Angle of view and light intensity has become the focus of current research.Go game is one of the most popular chess robots in the world,and the emergence of Alpha Go has also stimulated people's enthusiasm for artificial intelligence.At present,there is little research on the vision system of go robot.In fact,much of this research has focused on international chess.In addition,the visual system should be able to respond quickly to targets in the detection area.However,in practical application,the detection results may be affected by the specific environment of some checkerboard in practical application.The deep learningbased approach is computationally demanding and may not be the best solution.Whether or not there is an overfitting problem,the accuracy of deep learning models generally increases with the increase in model size,which inevitably slows down the processing speed.Therefore,this thesis proposes a go robot vision system based on machine vision algorithm research,the design of the visual system is used to calibrate and chess board testing identification method: first use watershed segmentation,simple geometric transformation and Hough transform algorithm to go robot vision system calibration,secondly using image correction,Hough transform,template matching to board piece realizes the robust target recognition,finally with the several popular convolutional neural networks are compared.The main contributions are summarized as follows:(1)Based on watershed segmentation algorithm,most of the current image segmentation algorithm can only segment the checkerboard,but not the area within the checkerboard boundary.In addition,the image is in a complex and changeable environment,and the traditional watershed segmentation algorithm is very prone to over-segmentation due to the direct impact of noise,which may lead to a large number of small segmentation areas in the subsequent segmentation process and results,affecting the final segmentation effect.Aiming at the above problems,this thesis proposes an improved watershed segmentation method,which can effectively solve the above problems and accurately locate the checkerboard boundary,which is very important for real-time application.(2)In this thesis,projection transform combined with Hough transform is proposed to detect corner points of checkerboard lattice,and a checkerboard reference model is constructed for corner point matching.Firstly,the checkerboard boundary obtained after dividing the watershed was used to calculate the four vertex coordinates of the checkerboard.Secondly,the complex background outside the checkerboard was removed by using the projection transformation.All the checkerboard lines in the checkerboard were accurately detected by using the Hough transformation and the positions of 361 checkerboard corner points were marked.Finally,a reference model is defined,and the predicted intersection of the midline of the reference image is obtained.(3)Finally,the input deformation image is corrected and the noise is removed from the corrected image.Secondly,the Hough transform is used to detect the empty lines on the board at the beginning of the game.Then use the template circle matching to locate the circle around the intersection of the checkerboard lines.Finally,for each intersection,calculate the average brightness value around it and determine a good threshold to identify the color of the pieces.We also design and construct several convolutional neural networks and compare them with the methods in this thesis,the superiority of the algorithm in this paper is verified.