基于多尺度注意力机制TransUNet的双目视觉定位与测量方法

打开文本图片集

关键词：双目视觉；TransUNet;关键点检测；注意力机制

中图分类号：TP391.41 文献标识码：A

doi：10.37188/OPE.20253316.2502 CSTR：32169.14.OPE.20253316.2502

Binocular vision localization and measurement method based on TransUNet with multi-scale attention mechanism

YANG Yu，XU Sixiang，ZHANG Mengquan，WU Duanzheng （College ofMechanical Engineering，Anhui University of Technology，Ma'anshan 243032，China） * Corresponding author， E -mail： xsxhust@ahut. edu. cn

Abstract：Aiming at the problems oflow detection eficiency of traditional binocular-vision feature-detection algorithms，as well as the insuficient attention to globally important features and the excessive parameter count of most network models，a method of continuous-casting-billet localization and measurement based on TransUNet binocular vision with a multiscale attntion mechanism was proposed. Firstly，left and right images of continuous-casting billets were colected with a calibrated parallel binocular camera to build a dataset. Subsequently，taking TransUNet as the backbone，an improved Transformer layer Was introduced to extract global context information； a Global Spatial Group Attention （GSGA） module was appended after every decoder block to enhance focus on globally salient features through a grouped multiscale attention mechanism; and a Convolutional Block Attention Module （CBAM） was inserted after each encoder-decoder skip connection and bilinear interpolation to boost key-point recognition by combining spatial and channel attention.Finally，3-D coordinate reconstruction and distance measurement were performed on the network’s key-point coordinates by leveraging binocular-vision principles. The experimental results show that compared with the Transformer model，the root-meansquare error and normalized error are reduced by 33.8% and 36.83% ，the number of parameters and floating-point operations are reduced by 10.58% and 8.21% ，and the single-batch inference time is shortened by 32.30% . In 3D ranging，the relative error of measurement reaches 0.137% ，which is significantly better than the traditional feature detection algorithm and meets the binocular vision localization and measurement requirements.

Key Words：binocular vision；TransUNet；keypoints detection；attention mechanism

1引言

现代钢铁冶金行业连铸坏经火焰切割，切割面上的熔化钢液在切割口的下边缘形成不规则且硬度大的毛刺，对钢材的表面质量和运输辊道造成不利影响。（剩余17786字）