CNN-based method for blotches and scratches detection in archived videos

In this work, we present a fully connected convolutional encoder-decoder for defects detection in archived video. The proposed method handles the detection of two of the most common archived video-related defects, namely blotches and scratches. It consists of two stages: (1) pixel-level classification and description of each video frame into defects pixels or not, by means of a novel CNN-based encoder-decoder architecture, and (2) spatio-temporal analysis to group and fine-tune the detections. For blotch detection, the learned features, extracted from an intermediate stage of the network, are used to evaluate the dissimilarity between the pre-selected regions in consecutive frames. For scratches detection, the morphology of scratches is used to eliminate false alarms. The experiments are performed on various video sequences suffering synthetic and real scratches and blotches. The results demonstrate the effectiveness of our approach and significant improvement against the most recent detectors. (C) 2019 Elsevier Inc. All rights reserved.