Image Processing Reference
In-Depth Information
6.2.4
Improved Temporal Image Signature Approach
The proposed approach [
30
] improves the TIS approach [
29
] by adding the face
features. The proposed approach presents how a combination of the Temporal
Image Signature approach (TIS) and face detection [
23
] will allow detecting
visually prominent features separately. In this section a brief and comprehensive
overview of the proposed approach is described.
Detecting salient areas within the scene is a challenging task. Researchers are
proposing different ideas and clues for detecting salient regions. This detection
becomes even more difficult while dealing with complex scenes. Factors such as
face detection can play an important clue in guiding attention to the area of interest.
The proposed approach provides useful information about the location of human
face where it is present in conjunction with detecting the salient regions in the
scene. This information could be a useful tool for detecting human gaze attention.
The idea of combining a face map with the saliency map is used in [
8
] where
Cerf et al. suggest combining a saliency model with the well-known Viola and
Jones matching algorithm [
37
] whereas Schauerte et al. [
5
] combine Modified
Cosine Transform (MCT) based face detection [
14
] with their saliency model.
Nevertheless in this paper, a method proposed in [
23
]
1
is used for the proof-of-
concept for face detection.
The proposed approach can be divided into the following main steps:
1. Frame division
2. Saliency detection
3. Face detection
4. Transformation and accumulation
Given a video clip as an input, frame division and saliency detection step are
performed in the same way as it was done in the case of TIS approach [
29
].
Face Detection In parallel to the salient regions detection, the detection of face is
done by applying face detection method [
23
] on each frame separately in XY
domain. After face detection, transformation of a detected face into a face map is
done using 2D Gaussian weighted function [
5
].
Transformation and Accumulation In the final step, provided the saliency infor-
mation in their respective planes, first salient information in the corresponding
planes (XT and YT) is accumulated by transformation back into the XY domain and
it is then fused with the face map. This results in a final map which is a combination
of saliency map and face map. Figure
6.5
shows the proposed system.
1
Matlab implementation of the face detection algorithm is available on
http://people.kyb.
