Research On Cell Culture Plate Detection And Grasping Method Based On Deep Learning

As a carrier to promote the "third revolution of biotechnology",synthetic biology has become a research hotspot in biological science,and has entered the fields of medical treatment,environment,energy,food,materials and so on.In the traditional synthetic biology laboratory,the detection and transport of cell culture plates are usually carried out manually,which is not only inefficient but also prone to errors.This situation urgently needs to be improved.Using vision guided robot instead of manual to detect and transport cell culture plate not only improves the experimental efficiency,but also reduces the error rate.However,the traditional target detection methods rely on the manual extractor,which has the problems of complex image processing,limited application scenarios and poor recognition robustness.In response to the above problems,this paper proposes an improved object detection method based on yolov5 s,which combines with robot vision to realize the detection,positioning and grasping of cell culture plate.The research contents of this paper are as follows:(1)The calibration and image registration of Eye-in-hand vision system were completed.By studying the imaging model of the camera,the internal and external parameters of the camera were obtained,and the hand eye calibration experiment of the vision system was completed.The RGB image and depth image of the D415 camera were aligned.The three-dimensional coordinates of the cell culture plate relative to the camera coordinate system are obtained,which lays the foundation for the subsequent realization of manipulator grasping.(2)Aiming at the problem of cell culture plate detection,the YOLOv5 s algorithm was improved.Firstly,deep separable convolution was used instead of conventional convolution in backbone network to make the model lighter.Secondly,K-means++clustering algorithm was used to select the anchor to make the prediction frame more accurate.Then the DIo U loss function was used to improve the convergence speed and accelerate the loss regression.Finally,Soft-NMS was used to optimize the candidate box,and convolution block attention mechanism is added to optimize the algorithm.(3)In the experimental environment,cell culture plate grabbing system was built.Then,the performance of the improved YOLOv5 s model and the accuracy of the cell culture plate grabbing system are analyzed.By comparing the image enhancement ability of SSR,MSR and MSRCR,MSRCR significantly optimized the image quality.In addition,the improved yolov5 s was used as the detection network of cell culture plate,and the training test was carried out.Compared with the original YOLOv5 s in terms of m AP,weight,time-consuming,etc.,the size of the improved YOLOv5 s model is reduced by 56.3%,and the detection accuracy is increased by 2.6%.Finally,the experiment of grasping cell culture plate was carried out.The average error rate of x,y and z axes was1.39%,2.41% and 0.29%,respectively,and the success rate of grasping was 84%.The experimental results show that the method of cell culture plate detection and grasping has good positioning accuracy and grasping performance,which has a certain reference value in practical application.