Abstract
γ-photon tomography technology, leveraging the high penetration of high-energy γ-photons and their electrical neutrality, has become a groundbreaking non-contact method for monitoring industrial flow fields. However, these fields typically appear as grayscale images. To enable parameter inversion, such as temperature and pressure mapping, these images must be colorized. Standard algorithms like DeOldify often produce "color gradient jumps" due to the probabilistic nature of γ-photon imaging. This paper introduces the Physical Enhancement Colorization Network (PECN), a U-Net+GAN framework that integrates multi-scale attention and physical perception augmentation. Using ResNet101 as the generator backbone and a discriminator that adaptively fuses global and local information, the PECN model enhances physical consistency. Validation on a SW60B turbojet engine simulation achieved PSNR, SSIM, FID, and MAE values of 32.5831, 0.8612, 17.8514, and 0.0191, respectively, outperforming the baseline DeOldify algorithm across all metrics.
Introduction
Monitoring the internal flow fields of industrial equipment, such as jet engines, is critical but difficult due to extreme thermal and pressure conditions. γ-photon tomography offers a non-contact solution by peering through these environments, yet its raw output is a grayscale probabilistic distribution of photon annihilation events. To utilize this data for engineering analysis, these images must be converted into color density maps representing physical properties. Standard colorization algorithms designed for natural photos often fail here, creating unnatural "color gradient jumps" that violate physical laws and render the data useless for precise analysis. The PECN model addresses this by prioritizing physical plausibility over mere visual aesthetics.
Methodology
The PECN model is built upon a Conditional Generative Adversarial Network (cGAN) framework with a U-Net generator. The architecture is designed to maintain structural integrity while accurately mapping grayscale values to physical color scales.
Physical Aware Enhancement (PAE) Module
The core innovation of the PECN model is the Physical Aware Enhancement (PAE) module, which is integrated into multiple levels of both the generator and discriminator. The PAE module consists of several sophisticated sub-components:
- Multi-scale Dilated Convolutions: Four parallel branches capture features at different scales, from fine jet boundaries to overall flow structures.
- Non-Local Blocks: These capture long-range dependencies, ensuring that color choices are physically consistent across the entire image.
- CBAM (Convolutional Block Attention Module): This helps the model focus on critical features and their spatial locations.
- Adaptive Fusion: A mechanism that decides the optimal balance between original structural features and newly generated color features.
Loss Function and Training
The model utilizes a hybrid loss function that combines Wasserstein distance for realism with VGG-based content loss for structural integrity. An adaptive weighting strategy dynamically shifts the focus from structural preservation to color saturation as training progresses, preventing early-stage distortion.
Data Generation
To overcome the lack of labeled real-world data, the researchers developed a high-fidelity synthetic pipeline. Computational Fluid Dynamics (CFD) simulations of a SW60B turbojet engine provided 42 ground-truth thermodynamic maps. These were converted into 3D phantoms and processed through the GATE Monte Carlo platform to simulate actual γ-photon detection. The resulting dataset was expanded via rotation, flipping, and noise addition to a total of 3,368 images (2,422 for training, 946 for validation).
Results and Discussion
The PECN model was tested against the baseline DeOldify and the state-of-the-art DDColor network, focusing on specific flow field features such as large-scale vortex wakes and horizontal wakes. It outperformed both across all quantitative metrics, achieving a PSNR of 32.5831 (a 4.54% improvement) and an SSIM of 0.8612 (a 2.82% improvement). Qualitatively, PECN images showed much smoother transitions from high-temperature (yellow-green) to low-temperature (blue) zones, matching the actual physical distribution of the engine's exhaust and eliminating the artifacts common in general-purpose models.
Conclusion
The PECN model successfully integrates physical constraints into a deep learning framework for γ-photon flow field colorization. By addressing the specific challenges of probabilistic imaging, it provides a reliable tool for industrial parameter inversion. Future work will focus on testing the model with real-world experimental data and optimizing the architecture for real-time monitoring in highly turbulent environments.
Summary updated