Research On Vision Recognition Of Image-guided Robot

It is of great significance to equip industrial robots with visual recognition system in automatic production to realize intelligent and active execution of industrial robots.In the production line of product forming,inspection,packaging,warehousing and other links,most of the product feature recognition and matching process adopts manual or semi-automatic operation process,high labor intensity and easy to make mistakes.The industrial robot combining image processing and automatic identification technology is beneficial to improve production efficiency and production comfort and reduce manual labor intensity.However,in the complex operating environment,image processing and automatic recognition technology is very difficult,because the recognition error will lead to machine misoperation.In order to overcome the difficulty of object recognition caused by image defects such as image pollution caused by hardware,background environment and target body itself,it is worth paying attention to the problem of image preprocessing properly and obtaining robust preprocessing algorithm.In this paper,with the Windows operating platform,combined with the C++,C# and the Python programming language implementation algorithm of image guided robot visual identification system design and software development,and focuses on the object contour size detection and localization,the character of image processing and automatic identification algorithm,to achieve the goal of image guided by three axis robot automatic grasping motion object.There are three parts in this paper:(1)research on object contour recognition and automatic positioning algorithm.First,two local exposure processing algorithms of "superposition" and adaptive parameter correction were developed by using automatic rectangle extraction algorithm,corrosion algorithm,HSV segmentation,blocking,gamma enhancement and minimum external rectangle,etc.Then,the image was segmented into basic image units by using the optimized edge detection algorithm.In view of whether there is object conglutination in the unit element of the image,the unit element is subdivided,and the contour information of the minimum unit element is extracted.The positioning error and size error between the source image contour and the template image contour obtained by the camera are calculated respectively through the contour matching algorithm.The results show that the arithmetic mean error of dime,pentagon and dollar are 0.17,0.01 and 0.23 pixels respectively,and the corresponding actual size error is 0.041,0.002 and 0.056 mm.(2)research on neural network recognition algorithm for character objects.First early the same image preprocessing step edge orientation,and position parameter and vertical projection method is used to change the character image segmentation as the smallest unit of image in yuan,then through data enhancement and image normalization to make 980 a training set,474 copies of validation set and test set,through sample training,the three layers of CNN model was established,to character recognition of characters,objects,and mark the characteristic information of the characters.In this paper,the Chinese chess character object is used for character recognition experiment.The results show that the CNN model of character recognition is as accurate as 99.09%.(3)visual recognition system design of image-guided robot.Firstly,a three-axis robot motion control platform was built.The X,Y and Z axes were respectively equipped with servo motors and their drives,and the motion control card was used to communicate with the upper computer.At the same time,the Z axis of the system is also equipped with an industrial camera and a grabber,through which the image of the working area can be acquired in real time.When the visual recognition system completes image capture and recognition,the starting position and target position coordinates of the motion of the manipulator are generated,and the motion instructions are sent to the motion execution unit through the motion control card to complete the object carrying action.The system test shows that,on the one hand,the absolute error is used to calculate the radius of the multi-face coin.The experimental results show that the edge error of the coin is less than 1 pixel,and the success rate of robot grasping is 100%.On the other hand,the CNN training model is used to test 14 Chinese chess characters.The experimental results show that the recognition rate of characters is 99.09%,and the success rate of chess capture is 100%.