Global-Edge Dual-Path Semantic Image Segmentation for Transparent Objects

Transparent object segmentation holds significant application value in intelligent manufacturing and service robotics. However, due to their inherent optical properties and susceptibility to complex background interference, conventional methods often suffer from low segmentation accuracy and boundary leakage. To address these challenges, this paper proposes a boundary segmentation module for edge enhancement of transparent objects. We implement this module within a semantic segmentation framework based on DeeplabV3+ architecture with ResNet50 as the backbone network, and design a dual-path learning strategy that integrates edge segmentation with holistic segmentation to enhance the model's capability in capturing detailed features of transparent objects. Experimental results on the public Trans10K dataset demonstrate the superior performance of our method, achieving a mean Intersection over Union (mIoU) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 8. 7 3 \%}$</tex> and Pixel Accuracy (PA) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{9 2. 8 3 \%}$</tex>. This results verifies that the proposed model effectively addresses key challenges in transparent object segmentation, providing a highprecision solution for industrial sorting and robotic visual perception tasks.