MIMO-Uformer: A Transformer-Based Image Deblurring Network for Vehicle Surveillance Scenarios

Area of Science:

• Computer Vision
• Deep Learning
• Image Restoration

Background:

• Motion blur significantly degrades surveillance image quality, hindering information retrieval.
• Convolutional Neural Networks (CNNs) have advanced image deblurring, but transformers offer superior deep representation learning.
• High computational complexity of standard transformers limits their application to high-resolution images.

Purpose of the Study:

• To propose an efficient transformer-based deblurring network, MIMO-Uformer, tailored for vehicle-surveillance scenarios.
• To address the challenge of local motion blur prevalent in surveillance footage.
• To enhance the performance of image deblurring by incorporating multi-scale image processing and specialized loss functions.

Main Methods:

• Developed MIMO-Uformer, integrating window-based multi-head self-attention (W-MSA) from Swin transformers into a multi-input, multi-output U-Net (MIMO-UNet) architecture.
• Employed multi-scale image processing within the MIMO-UNet framework to leverage diverse image resolutions.
• Introduced an Intersection over Patch (IoP) factor and a supervised morphological loss to specifically target and mitigate local motion blur.

Main Results:

• MIMO-Uformer demonstrated significant improvements in deblurring performance, particularly for local motion blur.
• Experiments on public (GOPRO) and self-established datasets showed notable gains in Peak Signal-to-Noise Ratio (PSNR).
• PSNR improvements of at least 0.21 dB on GOPRO and 0.74 dB on the self-established dataset were achieved compared to existing benchmarks.

Conclusions:

• MIMO-Uformer offers an efficient and effective solution for deblurring surveillance images, especially those affected by local motion blur.
• The proposed IoP factor and morphological loss enhance the network's ability to handle specific types of blur common in vehicle surveillance.
• The study validates the effectiveness of the proposed methods through extensive experiments, paving the way for improved surveillance systems.