Variant-Depth Neural Networks for Deblurring Traffic Images in Intelligent Transportation Systems

Intelligent transportation systems (ITS) with surveillance cameras capture traffic images or videos. However, images or videos in ITS often encounter blurs due to various reasons. Considering resource limitations, although recent technologies make progress in image-deblurring, there are still challenges in applying image-deblurring models in practical transportation systems: the model size and the running time. This work proposes an artful variant-depth network (VDN) to address the challenges. We design variant-depth sub-networks in a coarse-to-fine manner to improve the deblurring effect. We also adopt a new connection namely stack connection to connect all sub-networks to reduce the running time and model size while maintaining high deblurring quality. We evaluate the proposed VDN with the state-of-the-art (SOTA) methods on several typical datasets. Results on Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) show that the VDN outperforms SOTA image-deblurring methods. Furthermore, the VDN also has the shortest running time and the smallest model size.