Lightweight Model-based Intelligent Object Detection for Digital Twin in Infrastructure Construction Progress

Objection detection is a key aspect of digital twin enabled intelligent construction progress monitoring and management. However, the construction phase of power infrastructure engineering differs from conventional scenarios and is susceptible to various influencing factors, making it difficult to perform accurate object detection in such complex scenarios. This paper proposes a lightweight object detection framework that combines a lightweight neural network model with image enhancement algorithms and attention mechanisms, enabling accurate detection results with fewer model parameters. Firstly, to address issues of image detail loss and low-light images, the MSRCP algorithm and an image contrast enhancement algorithm based on an exposure fusion framework are used for image enhancement. A lightweight Mask R-CNN is then constructed, introducing depthwise separable convolution in place of convolution, significantly reducing the number of parameters. Additionally, a Convolutional Block Attention Module (CBAM) is incorporated into the backbone network of the model to enhance the network's feature extraction capability. Experimental results demonstrate that, on a construction dataset collected from power infrastructure engineering construction scenarios, the proposed method achieves a mAP of 86.11%.