改进U-Net 的光条纹分割算法

针对传统基于线结构光的视觉测量系统存在光条纹分割精度低的问题，提出了一种改进U-Net 的光条纹分割算法。改进算法使用VGG16 的卷积池化层代替U-Net 编码块中的卷积池化层，在U-Net 编-解码层间的跳连接中引入坐标注意力机制，在U-Net 编码块末端接入金字塔池化模块，采用Dice 函数和交叉熵函数的组合作为网络的损失函数，解决了光条纹占比失衡问题。基于线结构光测量原理，设计了工件尺寸测量系统。实验结果表明：改进U-Net 算法的平均像素准确度（mpa）为95. 61%，平均交并比（mIoU）为89. 73%，均高于其他对比算法；工件测量尺寸的绝对误差小于0. 1 mm，相对误差小于1%，重复精度小于0. 2%，满足工件的检测要求。

Aiming at the problem of low light stripe segmentation accuracy in traditional line-structured light-based visual measurement systems, an improved U-Net-based light stripe segmentation algorithm is proposed. The improved algorithm replaces the convolutional-pooling layers in the U-Net encoder blocks with the corresponding layers from VGG16, introduces a coordinate attention mechanism into the skip connections between the U-Net encoder and decoder layers, and attaches a pyramid pooling module to the end of the U-Net encoder blocks. A combination of Dice loss and cross-entropy loss is adopted as the network's loss function, which addresses the issue of imbalanced pixel proportion of light stripes. Based on the principle of line-structured light measurement, a workpiece size measurement system is designed. Experimental results show that the improved U-Net algorithm achieves a mean pixel accuracy (mpa) of 95.61% and a mean intersection over union (mIoU) of 89.73%, both higher than those of other comparative algorithms; the absolute error of the measured workpiece size is less than 0.1 mm, the relative error is less than 1%, and the repeat precision is less than 0.2%, which meets the workpiece detection requirements.

搜集汇总

背景与挑战

背景概述

该数据集是一篇关于改进U-Net算法用于光条纹分割的研究论文，通过引入多种技术提高了分割精度，实验结果显示其在多项指标上优于其他算法，并成功应用于工件尺寸测量。