End-to-End Image Compression With Segmentation Guided Dual Coding for Wind Turbines.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Transferring large, high-resolution images from wind turbine inspections is a significant bottleneck.
• Current methods struggle to balance image fidelity with efficient data transfer for defect detection.

Purpose of the Study:

• To develop an end-to-end deep learning framework for joint segmentation and dual-mode compression of wind turbine inspection images.
• To improve the efficiency of data transfer while preserving critical details for defect analysis.

Main Methods:

• A novel framework integrating a segmentation network (BU-Netv2+P) with CRF-regularized loss for precise blade localization.
• A hyperprior-based autoencoder for optimized lossy compression of non-critical image regions.
• An extended bits-back coder with hierarchical models for lossless compression of the region-of-interest (blade).

Main Results:

• The framework achieves superior compression performance and efficiency on a large-scale wind turbine dataset.
• It enables high-fidelity compression of blade regions and efficient background compression.
• Demonstrates parallelized dual-mode compression by reusing background-coded bits.

Conclusions:

• This is the first fully integrated learning-based region-of-interest (ROI) codec combining segmentation, lossy, and lossless compression.
• The proposed method offers a practical solution for automated wind turbine inspections by overcoming data transfer bottlenecks.
• Ensures that defect detection is not compromised by efficient image compression.