Development Of Image Segmentation System For Mixed-flow Panel Spraying Based On Deep Learning

The industrial spraying production field mainly uses three forms of manual operation,reciprocating machine cycle and robot spraying.With the development of the times,people are no longer satisfied with the completeness of the basic functions of the product,and put forward higher demands for unique appearance and design elements.Robot spraying is the best solution pursued by the current industry,but whether it is online programming or offline programming,or manual dragging and teaching spraying robots,they are cumbersome,inefficient,and unable to flexibly respond to the needs of diversified production.The problem is not applicable to small and medium-sized enterprises with small output and diverse products.Based on the above-mentioned pain points,this paper takes the sheet metal sheet spraying project of a robot company in Foshan as the background,and proposes a deep learning-based machine vision system that can realize the intelligent spraying requirements of any similar types of workpieces in a complex industrial environment.This article is divided into machine vision system and robot spraying system according to station design.The core machine vision system can be divided into four modules according to function: camera control and calibration,semantic segmentation training and prediction,path planning,robot communication and control;robot spraying system Including robot control cabinet,robot body and spray gun module.The research content of this paper includes: plan design and hardware selection according to the particularity of spraying environment,and verify the feasibility.Camera calibration is completed based on Zhang Zhengyou's checkerboard calibration method,combined with the nine-point calibration method,a fitting method is proposed to solve the depth Zc to improve the accuracy of hand-eye calibration;the image is preprocessed,a variety of traditional image segmentation techniques are used to verify the limitations,and data are collected for small and medium-sized enterprises Difficult features proposed to use U-Net neural network structure for the semantic segmentation task of sprayed workpieces;proposed a method of self-building small data sets and data enhancement,conducted network model training experiments and analyzed the experimental results of segmentation accuracy;proposed in combination with spraying process requirements A path planning method and complete coordinate transformation,and then combine the robot communication protocol to establish RS232 serial communication to complete data transmission and function control tasks.After completing the targeted tests of the above functions,the comprehensive project needs to use Python to develop industrial software programs based on Pycharm,OpenCV,and Pytorch,and to design system software interfaces based on Py Qt5;use a variety of workpieces to test software functions and robustness under different conditions,and to The error of the output result of the algorithm is analyzed.