Real-time Transparent Object Segmentation Based on Improved DeepLabv3+
Zhengguang Xu, Benshan Lai, Yuan Li, Tao Liu · 2021 China Automation Congress (CAC) · 2021
To solve the problem of inaccurate segmentation caused by the similarity of adjacent pixels between the transparent object and its background, we propose a real-time segmentation model for transparent object based on DeepLabV3+. We design a new feature extraction network for semantic segmentation on the basis of Darknet53 structure, and refactor the Atrous Spatial Pyramid Pooling module by using dense connection between different atrous convolution blocks to restore more detailed information. Compared with Deeplabv3+, our method improves accuracy in transparent object segmentation, effectively reducing the boundary pixel classification error, and our segmentation speed is 1.6 times that of the original DeepLabv3+, indicating that our method can be faster and more effective on the segmentation of transparent object in the context of natural experiments.