ResNet-50 Based Improved Model for Road Surface Monitoring: Enhanced Accuracy and Practical Applications

This paper proposes an improved model based on ResNet-50 for road surface monitoring tasks. The model has been improved in feature selection, downsampling, and feature fusion. Several popular classification models were selected based on a standard training dataset, and ResNet-50 was chosen as the foundation network model, taking into account the balance between accuracy and time complexity. The algorithm was then improved by adding attention mechanism, modifying downsampling methods, and adding transfer learning. Through experiments on a self-built dataset, the original ResNet-50 model was tested with an accuracy of 84.46%. After adding attention modules and improved downsampling methods in ResNet-50A model, the accuracy of the model increased to 88.4% and the loss decreased by nearly half. Finally, after adding transfer learning to the ResNet-50-A model, the accuracy of the ResNet-50-B model increased to 92.08% and the loss further decreased. However, the ultimate research result of this paper not only can improve accuracy but also can further expand the monitoring method, such as improving image classification to object detection, further improving the accuracy of practical applications. It can also be extended to practical applications beyond construction waste, without increasing the computational complexity, the focus of future research will be on improving practical applications.