Image deblurring by multi-scale modified U-Net using dilated convolution.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Urban traffic monitoring systems rely on license plate recognition, often hindered by motion blur from fast-moving vehicles.
• Deep learning offers potential for image deblurring to enhance vehicle information extraction.

Purpose of the Study:

• To develop an effective deep learning method for deblurring dynamic motion-blurred images.
• To improve vehicle recognition rates in urban traffic monitoring.

Main Methods:

• Proposed a multi-scale modified U-Net network incorporating dilated convolution.
• Implemented a variable scaling iterative strategy for adaptability to real-world blurred images.
• Utilized multi-scale architecture and dilated convolutions to enhance feature learning and receptive fields without increasing computational cost.

Main Results:

• The proposed method demonstrated favorable deblurring effects on both synthetic and real motion-blurred image datasets.
• Experimental comparisons showed the effectiveness of the new deblurring technique against existing methods.

Conclusions:

• The multi-scale modified U-Net with dilated convolution is an effective solution for deblurring motion-blurred images in traffic monitoring.
• The proposed method enhances image quality, leading to improved vehicle information retrieval and recognition accuracy.