Abstract: To improve the identification efficiency of concrete dam surface operation scenes, a new intelligent identification method (ResNet50-SEMSF) for typical scenes is developed. The collected monitoring video of the construction scenes is segmented into images, and their features－such as workers, machines, materials, environment, and other entity elements－are examined to define the typical scenes on a dam surface. With Residual Network 50 as the backbone network structure, a squeeze excitation attention mechanism is adopted to enhance the capability of expressing the key features of multi-target entity elements in the operation images. The down-sampling multi-scale features of an operation image are fused so as to retain its low-level features and high-level semantic information, enhance the model's capability of understanding the features at different levels, and overcome the difficulties in scale change and target deformation. With comparative analysis of the test results by other three convolutional neural network models, the Grad Class Activation Mapping visualisation method is used to illustrate the extent to which our new model focuses on information about the entity elements in the scene categories. The results show its recognition effect is significantly better than that of ResNet50, MobileNetV2 and VGG16 classical network models, characterising its feasibility and usefulness for concrete dam face operation in intelligent scene recognition and safety management.

Key words: concrete dam, dam surface operation, deep learning, attention mechanism, scene intelligent recognition