Rethinking deinterlacing for early interlaced videos

Abstract

With the rapid development of image restoration techniques, high-definition reconstruction of early videos has achieved impressive results. However, there are few studies about the interlacing artifacts that often appear in early videos and significantly affect visual perception. Traditional deinterlacing approaches are mainly focused on early interlacing scanning systems and thus cannot handle the complex and complicated artifacts in real-world early interlaced videos. Hence, this paper proposes a specific deinterlacing network (DIN), which is motivated by the traditional deinterlacing strategy. The proposed DIN consists of two stages, i.e., a cooperative vertical interpolation stage for split fields, and a merging stage that is applied to perceive movements and remove ghost artifacts. Experimental results demonstrate that the proposed method can effectively remove complex artifacts in early interlaced videos.

Multi-frame Joint Enhancement for Early Interlaced Videos

“…The proposed MFDIN is compared with traditional single field vertical interpolation (SFI) and recent DNN-based deinterlacing methods, i.e., DICNN [4] and DIN [5]. In addition, we also retrained a deeper single-frame model EDSR [3] with 16RBs and a SOTA multi-frame model EDVR [7].…”

“…In this paper, the degradation model of early interlaced videos is defined as I = f codec (f interlace (T odd , T even )) + n, where f codec denotes the video compression process and n represents noises. Existing jointly deinterlacing and denoise methods [5], [6] are based on single-frame reconstruction and cannot make full use of the temporal similarity between adjacent frames. However, interlaced frames are highly correlated with adjacent frames in time series, hence the performance of current single-frame de-interlacing networks [4]- [6] still has a lot of room for improvement.…”

“…However, the learning ability of shallow networks is limited, and complex artifacts in realworld interlaced videos cannot be handled well in this manner. Zhao et al [5] proposed a deeper deinterlacing network (DIN) for early videos by vertically up-sampling and merging two fields. As mentioned before, these DNN-based methods are all based on a single frame, which cannot sufficiently utilize inter-frame redundancy information.…”

“…The classical ResNet [39], which is commonly used in related lowlevel vision tasks [3] is selected as the backbone. Motivated by DIN [5], the odd and even fields of each frame are separately input for feature extraction, which can simulate the reverse process of manmade interlacing scanning mechanism. In this way, the network can perceive the difference among multiple fields, and further reduce the impact of interlacing artifacts during feature extraction.…”

“…In this way, the network can perceive the difference among multiple fields, and further reduce the impact of interlacing artifacts during feature extraction. Note that the resolution of the input fields is half of the original frame, we interpolate the reconstructed field features in the vertical direction to restore the original resolution by means of a vertical pixel shuffle (VPS) layer [5], [40], which is illustrated in Fig. 4.…”