Digital Image Semantic Segmentation Algorithms: A Survey 199
Where, E
1
(c
i
: x
i
) measures the probability that the pixel i is labelled c
i
under feature
x
i
, E
2
(c
i
: c
j
) measures the consistency of two connected pixels’ label.
In [11] a new high-order conditional random field was proposed. The model combined
the target detection results based on global shape features and the point-to-condition
model. Target detectors and pre-background segmentation algorithms were used to obtain
target regions in the image, and new high-level energy items were defined on the target
regions. The new high-order conditional random field model was a weighted mixed model
of high-order energy items and point-pair conditional random field models, its optimal
solution was the final semantic segmentation result of the image. The new high-order
energy term is defined as:
E
d
k
(x
d
k
) = −|x
d
k
|max(0, (1 − R)max(0, (C
d
k
− C
i
))) (7)
R =
N
d
k
R
t
|x
d
k
|
(8)
Where x
d
k
is a set of random flag variables corresponding to all pixels that make up
a single object area, C
t
is the threshold. By adjusting this value, the final recognition
accuracy rate can be controlled. Wang et al. [12] proposed an improved image segmen-
tation algorithm based on a robust high-order conditional random field model, according
to the given tag set, the maximum stream-minimum cut algorithm was applied to obtain
the local optimal solution, then the local optimal solution was used to modify the node’s
tag, and the extended algorithm was run on the unmarked nodes. At the same time, the
flow and edge of the graph were dynamically updated during each iteration, which would
make the time of each iteration decrease rapidly. The experimental results showed that
the convergence speed was faster on the same segmentation effect. The image semantic
segmentation algorithms based on conditional random field are shown in Table 2.
Table 2. Comparison of algorithms based on conditional random field (%)
Author Algorithm features Datasets
Segmentation
results
ZHANG[8]
CRF, dense features, high-order
potential energy
MSRC-21 75.8(mA)
ZHANG[9] CRF, Joint-boosting Algorithm MSRC-21 71.6(mA)
ZUO[10] CRF, Interactive Self-built dataset 95.3(mA)
MAO[11] CRF, high order energy items MSRC-21 72.2(PA)
WANG[12]
CRF, Maximum Flow -
Minimum Cut
MSRC-21 0.7s(time)
Chen et al. [13] proposed a new image semantic segmentation model in combination
with the underlying segmentation results. First, the corresponding underlying segmen-
tation image block was obtained by the histogram threshold and the K-means. Then
the high-level semantic information of the image was acquired by the word bag model.
Finally, the high-level semantic information was used in conjunction with the support
vector machine re-labels the image block to obtain the final image semantic segmenta-
tion result. In [14] an image semantic segmentation algorithm based on texture primitive
blocks was proposed. Firstly, texture primitive features were extracted, k-means and
k-d trees were used to get the image’s texture primitive block segmentation maps, and
then semantic mapping of texture primitive blocks was implemented by using the image
semantic learning and prediction methods based on support vector machine.
The two papers have similar ideas. Firstly, the image is subdivided and then the high-
level semantic information of the image is obtained. Then the support vector machine is