![](https://csdnimg.cn/release/download_crawler_static/18833579/bg3.jpg)
video analysis, which can be divided into two stages: (1) video structured analysis,
(2) human-face clustering. The former includes scene boundary detection and shot
detection. Subsequently, we perform face detection and face feature extraction and use
clustering algorithm for human-face clustering in the latter stage.
In our implementation, the method proposed in [9] is used to detect scene
boundaries for a movie. After the completion of scene boundaries detection, we use the
libfacedetection
1
engine for face detection, which can quickly and efficiently achieve
face detection. Based on it, we perform face feature extraction using SeetaFace
2
open-source face identification engine, in which the outputs of the 2048 nodes of the
FC2 layer in the VIPLFaceNet [8] are exploited as the feature of the input face. It is
worth mentioning that the model shipped in with the codes is trained with 1.4M face
images of 16K subjects including both Mongolians and Caucasians. Finally, the
hierarchical clustering method is used to cluster the human faces. Note, face occlusion,
angle changes, light change and other factors may lead to incorrect human face clus-
tering. Therefore, after the completion of a round of clustering, we select the faces,
clustering number of which is 1. Since these faces are most likely to be misclassified,
we use them to perform the next round of clustering by reducing the similarity
threshold, and cycle several times. The approach can effectively improve the clustering
accuracy.
In this paper, only the major characters in the video are analyzed, so after the
completion of face clustering, it is necessary to exclude faces with fewer clusters in the
results, and the remaining ones were judged as the major character. According to the
experimental results, the overall accuracy of face clustering reaches about 75%.
3 Construction of SRN
3.1 The Definition of SRN
A SRN is an undirected weighted graph expressed by
G ¼ C; E ; W
hi
ð1Þ
where C ¼ c
1
; c
2
; ...; c
n
fg
denotes the nodes in graph G, just the set of characters
that appear in the video. E denotes the set of edges in graph G, where e
ij
represents the
edge between c
i
and c
j
. The element in W represents the strength of the relationship
between characters.
In order to construct an SRN, it is necessary to quantify the “relationship” between
the characters, i.e. w
ij
. RoleNet [3] quantify characters’ relationship as the number of
co-occurrence between roles in the scene. However, this model can not accurately
reflect the relationship between characters, since the characters appear in the same
scene may have different importance, and should not be given the same weight.
1
https://github.com/ShiqiYu/libfacedetection
2
https://github.com/seetaface/SeetaFaceEngine
SRN: The Movie Character Relationship Analysis via Social Network 291