基于注意力机制和多尺度特征融合的氩花图像分割OA

In the bottom-blowing argon process of ladle metallurgy,the exposed area of the molten steel surface(referred to as the argon bubble region)serves as an indirect indicator of the argon gas flow rate.Accurate segmentation of the argon bubble region is crucial for optimizing the argon-blowing process.In view of this,the paper proposes a multi-branch argon bubble image segmentation algorithm based on attention mechanisms and multi-scale feature fusion,so as to enhance segmentation accuracy.The method is built upon the HRNet architecture and several improvements is introduced.Firstly,a combined Dice Focal Loss function is used to replace the conventional cross-entropy loss to eliminate the sample imbalance between the argon bubble and background regions.Secondly,depthwise separable convolutions are employed to reconstruct the residual modules of the backbone network,so as to significantly reduce both model parameters and computational cost.And then,a coordinate attention mechanism is integrated with minimal computational overhead to better get complex edge structures of the argon bubble regions.Finally,a parallel aggregation pyramid pooling module is designed to facilitate multi-scale information fusion and further improve segmentation accuracy.The experimental results show that the mIoU(mean intersection over union)of the proposed algorithm reaches 94.55%on a private dataset of a steel plant.It can be seen that the mIoU of the proposed algorithm increases by 1.62%and its accuracy reaches 99.31%compared with HRNet,so it can meet the requirements of argon bubble image segmentation in industrial production.