TF-ME:多尺度特征增强的透明物体分割网络  

TF-ME:Transparent Object Segmentation Network with Multi-Scale Feature Enhancement

作　　者：郭扬 邓晓军 肖世康 孙元昊 GUO Yang;DENG Xiaojun;XIAO Shikang;SUN Yuanhao(School of Computer,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007

出　　处：《湖南工业大学学报》2025年第5期58-66,共9页Journal of Hunan University of Technology

基　　金：湖南省自然科学基金资助项目(2024JJ7148)。

摘　　要：针对透明物体会继承来自背景的信息且传统卷积神经网络中感受野的限制等问题,提出了基于Transformer和多尺度特征增强的透明物体分割网络TF-ME。模型采用CNN结合Transformer的混合结构,在特征提取阶段,设计了多尺度特征融合模块,有效整合全局与局部信息,提升了模型对不同尺寸透明物体的分割效果;此外,对前馈神经网络进行了重新设计,增强了Transformer编码器的上下文理解能力。为验证所提算法的有效性,在Trans10K-v2数据集上进行了对比实验。实验结果表明,所提方法在11种透明物体分割中的ACC和MIoU值分别达到了94.68%和73.39%,相较于其他算法,该模型的性能明显提升。In view of the transparent objects inheriting information from the background and the limitation of receptive fields in traditional convolutional neural networks,a transparent object segmentation network TF-ME has been proposed based on Transformer and multi-scale feature enhancement.The model adopts a hybrid structure of CNN and Transformer.In the feature extraction stage,a multi-scale feature fusion module is designed to effectively integrate global and local information,thus improving the segmentation effect of the model on transparent objects of different sizes.In addition,the feedforward neural network is redesigned for an enhancement of the context understanding ability of the Transformer encoder,followed by comparative experiments conducted on the Trans10K-v2 dataset for a verification of the effectiveness of the proposed algorithm.The experimental results show that the proposed method achieves 94.68%ACC and 73.39%MIoU in 11 types of transparent object segmentation,respectively.Compared with other algorithms,the performance of the proposed model has been significantly improved.

关 键 词：透明物体 语义分割 TRANSFORMER 前馈神经网络 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]