Authors :

Mohammad Sharifi and Mohsen Zefreh

Address :

University of Isfahan — Khansar Campus

Khansar, Isfahan Province, Iran

Abstract :

The proposed Zebra Stripe Pattern inspired Convolutional Neural Network (ZSP-CNN) draws inspiration from the high-contrast, directional nature of zebra stripe patterns to enhance motion blur removal in traffic surveillance imagery. This design focuses on preserving structural edges and enhancing directional awareness in motion-degraded scenes. Existing motion deblurring methods, such as DeblurGAN and MPRNet, often fail to effectively handle directional blur, resulting in smoothed edges and a loss of critical details, particularly in dynamic traffic scenes where clarity is crucial for object identification and license plate recognition. To address these limitations, it introduces a Directional Feature-Aware Convolution (DFAC) framework, which leverages orientation-selective filters and attention mechanisms inspired by zebra stripe patterns. This approach allows the network to focus on dominant motion directions and better preserve structural integrity. The proposed ZSP-CNN architecture integrates DFAC layers within a multi-scale encoder-decoder framework, reinforced by a hybrid loss function that combines perceptual, PSNR, and directional edge losses. This enables high-fidelity deblurring while maintaining real-time inference speed. Experimental results on a motion-blur traffic dataset demonstrate that ZSP-CNN achieves superior performance, with a PSNR of 29.7 dB and an SSIM of 0.93, outperforming existing models.

Keywords :

Motion Blur Removal, Traffic Surveillance, Directional Convolution, Edge Preservation, Zebra Stripe Pattern, Convolutional Neural Network

DOI :