Deblur-YOLO: Real-Time Object Detection with Efficient Blind Motion Deblurring
Document Type
Conference Proceeding
Publication Date
7-18-2021
Abstract
Object detection has been a traditional yet open computer vision research field. In intensive studies, object detection models have achieved promising results regarding recognition accuracy and inference speed. However, previous state-of-the-art algorithms fail to operate at blurry images. In this work, we propose Deblur-YOLO, an efficient, YOLO-based and detection-driven approach robust to motion blur photographs. We introduce a generative adversarial network with a dilated feature pyramid generator, a pair of multi-scale discriminators with spectral normalization, and a detection discriminator. We design a new image quality metric called Smooth Peak Signal-to-Noise Ratio (SPSNR) for measuring the smoothness of the reconstructed image. Empirical studies on benchmark datasets demonstrate Deblur-YOLO's superiority. On COCO 2014, Set 5 and Setl4, Deblur-YOLO achieves leading results for parameters, deblurring time, PSNR, SPSNR and SSIM. We also visually display the excellence of our deblurring performance to competing models.
Publication Title
Proceedings of the International Joint Conference on Neural Networks
DOI
10.1109/IJCNN52387.2021.9534352
Recommended Citation
Zheng, Shen; Wu, Yuxiong; Jiang, Shiyu; Lu, Changjie; and Gupta, Gaurav, "Deblur-YOLO: Real-Time Object Detection with Efficient Blind Motion Deblurring" (2021). Kean Publications. 943.
https://digitalcommons.kean.edu/keanpublications/943