Coarse-to-fine two-stage semantic video carving approach in digital forensics

Abstract Video (e.g. CCTV) plays a crucial role in digital forensics. Existing video carving methods either assume the existence of file system information or resort to the impractical exhaustive matching for all pairs of fragments. In this paper, a coarse-to-fine two-stage semantic video carving approach is proposed to improve the efficiency and precision of content-based video carving. The proposed approach introduces a perceptual grouping stage to quickly group video fragments first based on the structural similarity of the fragments, followed by a precise stitching stage which sorts the fragments within each group depending on the pixel-level content of the fragments to reconstruct each original video file. The proposed approach can reduce the computational complexity and achieve a high carving precision since the two complementary methods used in two stages focus on different-scale features of video content. Experimental results based on a YouTube-8M video clip dataset show that the overall carving precision of the proposed approach is very high (e.g. 97.2% even when the number of mixed video fragments is increased to 288, which are from 36 video files with a fragmentation degree of 8). The overall carving time is 326.22 seconds, about 10 times lower than that of our previous optical flow-based approach.