基于改进DeepLabv3+网络的氩花图像语义分割

Semantic segmentation of argon-flower images based on improved DeepLabv3+ network

钢包底吹氩过程中钢液表面裸露区域(即氩花)的面积可以间接反映吹入钢包中的氩气量。为了准确识别出钢水表面图像中的氩花区域,本文提出一种基于改进DeepLabv3网络的图像语义分割方法。该方法以DeepLabv3网络为基础,采用MobileNetV2作为主干特征提取网络,以降低网络的参数量和计算量;同时将原来的交叉熵损失函数替换成Focal Loss损失函数,以解决正/负样本不平衡和难/易分类样本不平衡的问题;最后在网络结构中添加通道注意力机制来提高语义分割精度。以生产现场采集的图像数据为对象进行实验,结果表明,与原始DeepLabv3相比,本文网络模型的参数量和计算量降低了约92.3%,平均交并比提升了0.82个百分点,达到92.4%,帧率提高了23.40%。

In the process of argon blowing from the bottom of a ladle, the area of exposed surface of the molten steel, i.e. argon flower, can indirectly reflect the amount of argon blown into the ladle. In order to accurately identify the argon flower area in the molten steel surface image, this paper proposes a semantic segmentation method based on improved DeepLabv3+ network. The method uses DeepLabv3+ as the basic architecture, and MobileNetV2 serves as the main feature extraction network to reduce the parameters and computation amount. At the same time, the original cross-entropy loss function is replaced by Focal Loss function to deal with the imbalance of positive/negative samples and hard/easy samples. Finally, a channel attention mechanism is integrated into the network structure to improve the accuracy of semantic segmentation. Experimental results on the image data collected at the production site show that, compared with the original DeepLabv3+, the parameters and computation amount of the proposed network model are reduced by about 92.3%, the mean intersection over union(MIoU) is increased by 0.82 percentage point to reach 92.4%, and the frame rate is increased by 23.40%.