An Efficient Deep Learning-Based High-Definition Image Compressed Sensing Framework for Large-Scene Construction Site Monitoring

Area of Science:

• Computer Vision
• Signal Processing
• Machine Learning

Background:

• High-definition (HD) images are crucial for construction site monitoring.
• Transmitting HD images from sites with poor networks and limited computing power is challenging.
• Existing deep learning methods for compressed sensing struggle with efficiency and accuracy for large-scale construction sites.

Purpose of the Study:

• To develop an efficient deep learning framework for compressed sensing and reconstruction of HD monitoring images from large-scale construction sites.
• To address the limitations of current methods in terms of memory usage, computational cost, and recovery performance.

Main Methods:

• Proposed an efficient deep learning-based high-definition image compressed sensing framework (EHDCS-Net).
• The framework comprises sampling, initial recovery, deep recovery body, and recovery head subnets, utilizing convolutional, downsampling, and pixelshuffle layers.
• Incorporated nonlinear transformations on downscaled feature maps and an efficient channel attention (ECA) module to reduce computational load and enhance reconstruction.

Main Results:

• The EHDCS-Net framework demonstrated reduced memory occupation and floating point operations (FLOPs).
• Achieved superior reconstruction accuracy compared to state-of-the-art methods.
• Showcased faster recovery speed in experiments on real hydraulic engineering megaproject data.

Conclusions:

• The EHDCS-Net framework offers an effective solution for compressed sensing of HD images in challenging construction site environments.
• The method balances efficiency (memory, computation, speed) with high reconstruction accuracy.
• This work advances the application of deep learning for image monitoring in infrastructure projects.