An efficient approach for image de-fencing based on conditional generative adversarial network

Abstract

Automated image de-fencing is an important area of computer vision that deals with the problem of virtually removing fence structures, if any, from images and produce aesthetically pleasing images without the fence structures. Unlike most of the previous de-fencing approaches that employ a two-stage process of fence mask detection followed by image inpainting, here we present a single-stage end-to-end conditional generative adversarial network-based de-fencing model that takes as input a fenced image and produces the corresponding de-fenced image in only 16 ms. The proposed network has been trained using an extensive dataset of fenced and ground-truth de-fenced image pairs by employing a combination of adversarial loss, L1 loss, perceptual loss, and estimated fence mask loss till convergence. The experimental results shows that our approach is capable of successfully handling images with even broken, irregular, and occluded fence structures. Qualitative and quantitative comparative study with previous de-fencing methods also show that our approach outperforms these existing techniques in terms of both response time and quality of de-fencing. © 2022, The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature.