Abstract
In Pressure-Sensitive Paint (PSP) imaging, deblurring is a critical step due to factors like prolonged camera exposure times and high model velocities, which frequently result in significant motion blur. Traditional deblurring methods often fall short in handling the non-uniform nature of this blur. This paper introduces the Deformable Fusion Deblurring Network (DFDNet), a novel deep learning architecture that leverages deformable attention and wavelet-based processing to restore high-fidelity pressure maps with minimal parameter overhead. By combining these techniques, DFDNet achieves state-of-the-art performance in both visual restoration and scientific measurement accuracy.
Introduction: The Challenge of PSP Imaging
Pressure-Sensitive Paint (PSP) is a cornerstone of modern aerodynamic testing, providing a non-contact method to visualize surface pressure distributions on aircraft models within wind tunnels. The technology relies on the oxygen-quenching of fluorescent molecules; however, capturing these subtle optical shifts requires relatively long exposure times. When coupled with the high-speed vibrations and camera shake inherent in wind tunnel environments, the resulting images suffer from complex, non-uniform motion blur.
Standard deblurring techniques typically assume a uniform blur kernel across the entire image. In contrast, PSP images of complex geometries exhibit varying blur amplitudes and directions. Conventional Convolutional Neural Networks (CNNs) also struggle in this context, as their rigid kernels cannot adapt to the spatially varying nature of the distortion. This leads to significant errors when the deblurred images are converted into quantitative pressure data, potentially compromising aerodynamic analysis.
Proposed Method: DFDNet
To address these limitations, we propose the Deformable Fusion Deblurring Network (DFDNet). DFDNet is built upon an encoder-decoder framework and utilizes several innovative modules to enhance restoration quality while maintaining extreme parameter efficiency.
Deformable Ack Attention Module (DAAM)
The core of DFDNet is the DAAM, which replaces standard rigid convolutions with variable kernels. By integrating attention mechanisms—reimagined as efficient 1x1 convolutions—with deformable convolutions, DAAM can adapt its receptive field to the specific shape and scale of the blur in different image regions. This allows the model to capture dynamic motion amplitudes without the prohibitive computational cost of standard Vision Transformers ($O(N^2)$ complexity).
Residual Shortcut Convolution (RSCBlock)
To further optimize the network, we introduce the RSCBlock. This module replaces traditional ResBlocks and is specifically designed to reduce spatial and channel redundancy. By streamlining feature extraction, the RSCBlock allows the network to achieve high performance with a significantly lower parameter count than traditional deep architectures.
Wavelet Downsampling and Dynamic Upsampling
Precise scientific measurements require the preservation of fine spatial details. Standard pooling operations often discard critical pixel-level information. DFDNet employs Haar wavelet downsampling (wt), which decomposes spatial information into the channel dimension, ensuring no data is lost during the encoding phase. For the decoder, dynamic point-based upsampling (dysample) is used to reconstruct high-resolution images with superior fidelity.
Dual-Network Training Framework
Deblurring is an inherently ill-posed problem. To stabilize training and ensure physical consistency, we employ a dual-network framework. A primary deblurring network is paired with a secondary re-blurring network. By enforcing a consistency loss—where the re-blurred version of the restored image must match the original blurry input—the model is prevented from introducing artifacts or "hallucinating" incorrect details.
Experimental Results
The DFDNet was evaluated on both specialized aerodynamic datasets and general-purpose public benchmarks.
Performance on PSP Data (FY Dataset)
Using the FY Dataset, which contains real-world PSP images from the China Aerodynamics Research and Development Center, DFDNet achieved a Peak Signal-to-Noise Ratio (PSNR) of 33.52 and a Structural Similarity Index (SSIM) of 0.949. Crucially, the Mean Relative Error (MRE) for pressure reconstruction was reduced to 0.0907, demonstrating the model's readiness for practical aerodynamic analysis.
Efficiency and Generalization
On the public GOPRO benchmark, DFDNet achieved a PSNR of 34.01 using only 5.29 million parameters. This is an order of magnitude more efficient than competing state-of-the-art models like NAFNet (67.9M) or UFPNet (80.3M). Furthermore, tests on the HIDE dataset confirmed the model's robust generalization capabilities across different types of motion blur without specific fine-tuning.
Conclusion
DFDNet represents a significant advancement in scientific image restoration. By combining deformable attention with wavelet-based spatial preservation, it provides a highly efficient and accurate solution for deblurring PSP images. While the current implementation involves high computational complexity per parameter, its superior accuracy in pressure measurement makes it a vital tool for the future of aerospace wind tunnel testing. Future work will focus on optimizing inference speeds and exploring domain adaptation for broader imaging applications.
Summary updated