【摘要】 水雾特征提取与及时检测对于提前预警铝加工（深井铸造）工艺容易产生的铝液泄漏爆炸事故有重要意义。为解决水雾检测标注数据集成本高、检测精度低、实时性不够等问题，提出了一种基于改进的DeeplabV3+和迁移学习的水雾分割模型。该模型采用MobilenetV2网络模型代替原算法的特征提取网络，有效降低模型参数量，显著提升模型的运行速度；引入迁移学习方法，解决样本不足造成的模型过拟合、分割性能不佳等问题；通过引入CBAM注意力机制，充分利用上下文信息，加强对水雾特征的学习。实验结果表明，改进的DeeplabV3+模型在水雾分割任务中取得了较好的表现，其中，MIoU达到了91.98%，水雾交并比达到了85.29%，较原始的网络模型提高了2.88个百分点，模型参数量大大减少，处理速度显著提高。更多还原

【Abstract】 The extraction and timely detection of water mist characteristics are of great significance for early warning of molten aluminum leakage and explosion accidents that are prone to occur in aluminum processing(deep well casting) processes. In order to solve the problems of high cost of water mist detection annotation data set,low detection accuracy,and insufficient real-time performance,a water mist segmentation model based on improved DeeplabV3+and transfer learning is proposed. This model uses the MobilenetV2 network model to replace the feature extraction network of the original algorithm,which effectively reduces the number of model parameters and significantly improves the running speed of the model;it introduces the transfer learning method to solve problems such as model overfitting and poor segmentation performance caused by insufficient samples;By introducing the CBAM attention mechanism,we make full use of contextual information and enhance the learning of water mist features. The experimental results show that the improved DeeplabV3+model achieved good performance in the water mist segmentation task. Among them,the MIoU reached 91. 98%,and the water mist intersection ratio reached 85. 29%,which is 2. 88 percentage points higher than the original network model. The number of model parameters is greatly reduced and the processing speed is significantly improved.更多还原