A Visual Sorting System For Recyclable Waste Based On Deep Learning

With the development of China’s economy and the substantial improvement of comprehensive national strength,the field of artificial intelligence and robotics has also achieved rapid development.The traditional mechanical arm grasping is mostly completed by teaching,which has certain limitations and strict requirements on grasping position.Artificial intelligence technology has brought great convenience to people in the service industry.Speech recognition reduces the complexity of text communication,human-computer interaction enhances user experience and brings convenience to people’s lives,and home intelligent robots improve people’s happiness index.With the demand of industrial development,artificial intelligence technology has also been developed in the industrial field.People can control the operation of machines and complete different tasks through speech recognition technology.Through deep learning visual inspection technology,industrial robots can have "eyes",have the same ability of recognition and discrimination as human beings,and can adapt to changes in the environment.In this paper,a recyclable garbage sorting system based on deep learning is designed for bottle-shaped objects in recyclable garbage,which replaces the traditional sorting method to liberate workers from the assembly line.The main research work is as follows:(1)Introduce the related technologies and development status of target detection and sorting robots.The whole sorting system is designed and determined by studying and consulting relevant materials in the early stage.The whole system is composed of multiple modules,including vision acquisition module,target detection module,robot body control,end-effector grasping module and so on.And complete the selection of the hardware equipment required in the system,the acquisition and processing of the camera connection images,the learning and application of the robot development software,and the network of target detection algorithms to build the communication between the modules.(2)Introduce the principle of binocular vision imaging distance measurement.According to the experimental environment and requirements,complete the fixing of binocular camera,complete the construction of the visual platform,and complete the measurement of the distance between the target object on the experimental platform and the lens by writing code based on the principle of ranging.According to the relationship between the image acquisition module,the target detection module and the robot grasping module,the overall layout of the system and the construction of the workbench are designed and completed.By using binocular camera to capture the calibration board information at different angles,the MATLAB software toolkit is used to complete the calibration of the camera parameters and the solution of the rotation matrix and translation matrix between binocular cameras,and then the calibration results are analyzed and processed.Multiple target pictures are taken,and the external parameter matrix is calculated by reading the coordinates of different target positions of the demonstrator,so as to realize the conversion of camera coordinate system to robot base coordinate system.(3)The deep learning target detection algorithm YOLOV4 is used to train the model to realize the positioning and classification of target objects.Firstly use the Pytorch deep learning framework to complete the construction of the overall network model according to the idea of the target detection algorithm,make and label the data set according to the pictures obtained by yourself,analyze the network structure and detection principle of the target detection algorithm,and combine your own data set to improve the network And optimization,using the k-means dimensional clustering optimization algorithm for the network framework to perform clustering analysis on the target frame in the data set,and optimize the setting of the candidate frame.A Mosaic data enhancement method is designed to splice multiple images,so that multiple images could be input to the network at one time for training,so as to obtain a better model and improve the detection of small targets.(4)Complete the robot arm motion planning and experimental verification,build the industrial robot visual sorting system,and apply the recognition and positioning algorithm to the system.Through the target detection network to realize the identification and positioning of the garbage,the position information and category information are sent to the robot control system through the socket communication to realize the grasping of the robot arm.The recognition rate and the sorting success rate for recyclable bottle-shaped garbage are 94%and 86%,respectively.The experimental results show that the recyclable garbage sorting system based on deep learning can effectively identify and locate the target object,and complete the grabbing by the robotic arm,which verifies the feasibility of the sorting system.