ResNet50-based classification of footwear in nuclear power plants surveillance images

In nuclear power plants, the correct wearing of work boots in accordance with regulations is crucial to protect the health of staff from the effects of radiation. Effective and accurate detection of employees wearing standardized shoes in nuclear power plants is imperative. Deep learning methods have the capability to extract image features and provide accurate image classification results. To achieve the automated classification of footwear, we devised a deep learning-based classification network. In this work, a footwear classification dataset is conducted from the human detection results of surveillance videos in a nuclear power plant. Using this dataset, we trained a classification model based on ResNet-50 to classify the dataset into five categories, including false detection, indeterminate, work boots, shoe covers and ordinary shoes. To mitigate the impact of class imbalance and avoid overfitting, training strategies including label smoothing and mixup data augmentation are employed during training. The final classification accuracy reached 96.62%.