Image Segmentation Algorithm in Complex Environment Based on Improved SOLOV2

Previous industrial robots combined with vision systems in target grasping tasks have high requirements for the task environment. Once the vision system is confronted with a target object in a complex environment such as stacking, tilting and occlusion, it will be difficult to provide the robot cont...

Full description

Saved in:
Bibliographic Details
Published in2023 12th International Conference of Information and Communication Technology (ICTech) pp. 581 - 585
Main Authors Zhou, RenJian, Zheng, LiaoMo, Ren, ChuLan, Wang, ShiYu
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.04.2023
Subjects
activation functions
Adaptation models
descriptor
Feature extraction
feature fusion
Image segmentation
Machine vision
Production systems
Robot control
SOLOv2
Stacking
vision systems
Online AccessGet full text

Cover

More Information
Summary:Previous industrial robots combined with vision systems in target grasping tasks have high requirements for the task environment. Once the vision system is confronted with a target object in a complex environment such as stacking, tilting and occlusion, it will be difficult to provide the robot control system with the correct target pose information due to the lack of target feature information, which is not conducive to the dynamic configuration capability of the production system. This study proposes an industrial target image segmentation algorithm based on improved SOLOv2 to solve the problem of low accuracy of industrial target image segmentation caused by small targets as well as occlusion. Building upon the SOLOv2 model, the optimization of fusing multi-layer features within the mask feature branch is deaigned to integrate feature information from the deepest to the shallowest layers, in a layer-by-layer manner; the Mish activation function is selected in the upsampling process to enhance the model's ability to extract small target features in images while improving the generalization ability and convergence speed of the model; to mitigate the impact of occlusion on image segmentation, the weights of the loss function are adaptively adjusted based on the degree of occlusion. This allows for a reduction in the negative influence of occlusion on image segmentation, as the loss function is optimezed to account for varying degrees of occlusion.
DOI:10.1109/ICTech58362.2023.00112