Wire melted mark metallographic image recognition and classification based on semantic segmentation

The recognition and classification of wire melted marks is crucial in modern fire investigation. The existing technology mainly uses physical or chemical methods to deal with wire melted marks and draws conclusions through manual observation, or manually extracts the features and train classification model, both of which consume excessive manpower and resources. The research on automatic feature extraction and recognition of wire weld marks by artificial intelligence technology is still blank. Based on the data set of wire melted mark metallographic images provided by a city fire research institute, we proposed an algorithm to recognize the type of wire melted mark metallographic images based on artificial intelligence which can help fire fighters efficiently speculate the cause of fire. In the algorithm, the TransUnet network is used to segment the melted zone by semantic segmentation to extract the melted zone containing the main features, and the mIOU reaches 92.2%. Then, the features of wire melted mark are extracted based on the melted zone image. Finally, XGBoost is used for feature modeling for classification. The F1 Score of model is 82.9%.