Unsupervised Video Summaries Using Multiple Features and Image Quality

Video summarisation with visual and semantic cues

“…These methods take advantage of certain types of features related to interestingness, therefore, fully representing interesting segments is difficult. Recently, numerous methods [8, 9, 11, 12] have combined a variety of high‐correlation features with interestingness for a summary. Gygli et al [11] combined the aesthetics/quality, spatio‐temperal saliency, faces/person, and landmark detection to compute the interestingness score.…”

“…Gygli et al [11] combined the aesthetics/quality, spatio‐temperal saliency, faces/person, and landmark detection to compute the interestingness score. Hu et al [8] considered image quality and incorporated other low‐level features as bases for evaluating the importance of video frames. These approaches perform better than methods based on a single feature, but they still do not consider semantic information.…”

“…A good video summary should satisfy at least two points: it extracts the most interesting segments of a video [7] and captures the theme of a video. Many existing video summarisation approaches [1, 4, 8–12] only focus on satisfying the visual interestingness of viewers. Some of these approaches based on one feature, such as visual attention [4, 10] and colour feature [1].…”

“…Some of these approaches based on one feature, such as visual attention [4, 10] and colour feature [1]. Other approaches [8, 9, 11, 12] fuse various visual features to extract important segments for summaries. Although these methods can generate video summaries, they disregard whether the extracted summary can accurately convey the theme of a video from the semantic level.…”

Phân lớp ảnh dựa trên tổ hợp đa đặc trưng

“…Khái niệm 4: (Đa đặc trưng) [7] Đa đặc trưng (multiple features) là một bộ gồm nhiều vector đặc trưng biểu diễn cho một ảnh hay một số ảnh con khác nhau của một ảnh đầu vào. Mỗi vector đặc trưng của bộ này được tạo ra từ một kĩ thuật trích chọn riêng hoặc là vector đại diện cho một ảnh con của ảnh đầu vào.…”