【摘要】 在KR铁水脱硫工艺中，高效的扒渣作业对提升最终钢材的纯净度至关重要。然而，由于铁水、炉渣与罐壁的视觉特征高度相似，自动化扒渣过程常面临液面识别困难、分割精度低等问题，严重制约了扒渣的效率与质量。针对上述挑战，本文提出一种改进的轻量化DeepLabV3+语义分割网络。该网络采用MobileNetV2替代原始的Xception主干网络，在降低模型参数量的同时提升运算效率；设计并构建了融合深度可分离卷积与通道注意力模块（CAM）的CAM-ASPP模块，实现了特征提取的轻量化与增强；此外，在网络中引入卷积块注意力模块（CBAM），进一步强化模型对铁水与炉渣等关键区域特征的辨识与表达能力。基于工业现场采集的数据集进行验证，结果表明，本文模型的平均交并比（MIoU）与平均像素准确率（MPA）分别达到92.86%和96.38%，较原始模型分别提升了1.18和1.01个百分点。此外，该模型内存占用仅为19.59 MB，相较于原始模型大幅降低了93.96%，这为实现高精度、轻量化的实时扒渣视觉检测提供了有效方案。更多还原

【Abstract】 In the KR hot metal desulfurization process,efficient slag removal is critical for improving the purity of the final steel product. However,the automated slag removal process is often hampered by the high visual similarity among molten iron,slag,and the ladle wall. This leads to difficulties in accurate liquid level detection and low segmentation accuracy,which severely constrain the efficiency and quality of slag removal. To address these challenges,this paper proposed an improved lightweight DeepLabV3+ semantic segmentation network. The proposed network utilized MobileNetV2 to replace the original Xception backbone,thereby reducing the number of model parameters while enhancing computational efficiency. Additionally,a CAM-ASPP module was designed and constructed by integrating depthwise separable convolutions with a Channel Attention Module(CAM)to achieve lightweight and enhanced feature extraction. Furthermore,a Convolutional Block Attention Module(CBAM)was incorporated into the network to further strengthen the model’s capability to identify and represent features of key regions,such as molten iron and slag. Validation on a dataset collected from an industrial site shows that the proposed model achieves a Mean Intersection over Union(MIoU)of 92.86% and a Mean Pixel Accuracy(MPA)of 96.38%,representing improvements of 1.18 and 1.01 percentage points,respectively,over the original model. Moreover,the model’s memory footprint is only 19.59 MB,a reduction of 93.96% compared to the original model,providing an effective solution for high-precision,lightweight,and real-time visual detection in slag removal operations.更多还原