Unsupervised Deblurring Algorithm Based on Deep Image Prior for Vehicle Detection Systems
Jiahao Chang, Peng Tang, Zirou Jiang, Zhentao Wang, Zhifang Wu, Yuewen Sun · IEEE Transactions on Nuclear Science · 2025
Radiation imaging vehicle inspection systems are widely employed in public security applications. However, at high inspection speeds, the resulting radiation images often become blurry and noisy, compromising the identification of contraband. Recently, deep image prior (DIP) algorithms and their variants, which do not require dataset pre-training, have demonstrated significant potential in addressing imaging inverse problems. Despite their promise, DIP methods are prone to overfitting when the number of training iterations surpasses a critical threshold, leading to a notable decline in image quality. To address this limitation, we propose an enhanced DIP framework that integrates total variation (TV) regularization, resulting in a dataset-free non-blind deblurring method named UA-DIP. This approach utilizes a flexible alternating direction method of multipliers (ADMM) to solve the optimization problem by constraining the solution space through iterative steps. Experimental results demonstrate that UA-DIP outperforms several classical algorithms, achieving notable improvements in evaluation metrics and delivering visually convincing deblurring results. Furthermore, the method achieves excellent performance in restoring real radiation images, underscoring its effectiveness in the field of radiation imaging.