Digital Signal Processing Reference
In-Depth Information
Skin-color
Filtering
Rejection
Cascade
AdaBoost Classifier
Face
…
Face
Non-Face
Non-Face
Non-Face
Non-Face
Non-Face
Fig. 1.12
Face detection cascade in [
62
]
representation of the shape of objects within a general class (such as horse images),
and then use this information to segment novel images. An interesting trend can be
observed, which aims to segment a collection of unlabeled images while exploiting
automatically discovered appearance patterns shared between them [
61
].
1.4
Challenges
In this section, we identify serious challenges that remain despite over ten years of
progress in image/video segmentation.
The first challenge in image segmentation is the semantic gap, which means how
to bridge the semantic gap between low-level features and high-level semantic effec-
tively. As stated in the last section, given a similarity matrix (e.g., intensity based),
many algorithms can be used to implement data clustering. But it is still difficult to
ensure the clustering result with meaningful partitions. For good segmentation, both
spatial and semantic relations are required. The demand is significantly urgent for
image analysis and scene understanding. Thus far, a range of techniques have been
developed to achieve 'good' segmentation by incorporating more semantic informa-
tion into the clustering process.
The second challenge is to yield accurate segmentation for images. Although
image segmentation is application oriented, it is not necessary to provide accurate
segmentation mask for some applications. However, accurate segmentation is par-
ticularly interesting in a lot of fields. Many are not willing to accept the coarse
segmentation results that may be obtained by the existing approaches. More specif-
ically, a robust and accurate segmentation will lead to the distinct improvement for
many applications such as content based video coding (e.g., MPEG-4). However, it
is still a challenging task to extract accurate object mask from the image/video be-
cause of the variations of brightness, lighting, view, and other complex backgrounds.
Despite the advances in video segmentation, it remains difficult to segment ob-
jects of interest in real time, which can be regarded as the third challenge. Unlike
the static image segmentation, it seems that temporal features such as motion that
