Image recognition technology based on neural network in robot vision system. (2nd December 2021)

• Record Type: Journal Article
• Authors: He, Yinggang
• Is Part Of: International journal of grid and utility computing. Volume 12:Number 4(2021)
• Page Start: 415
• Page End: 424
• Publication Date: 2021-12-02
• Subjects: neural network -- image recognition -- machine vision -- recognition system; Electronic data processing -- Distributed processing -- Periodicals; Electronic commerce -- Management -- Computer programs -- Periodicals; 004.605
• Languages: English
• ISSNs: 1741-847X

Abstract

This research uses CamVid training decoder to train the model, then fine tune the parameters on the collected data, label the manually collected data with LabetMe annotation tool, and cross verify the image and scene with neural network algorithm and image recognition principle technology. After five training cycles, the neural network in this study can achieve more than 90% recognition accuracy, and achieve convergence after storing about 10 cycles. Finally, the recognition accuracy in the test data set can reach more than 95%. In the range of robot vision recognition, the maximum measurement deviation is only 2.54 cm and the error is less than 2%. It can be concluded that the method used in this study has fast convergence speed, high recognition accuracy, small error, and good practicability and effectiveness.