[SM
1ina
and more
ition, and
attention
le. In our
e placed.
signs. In
; saliency
d into two
ase, they
transform
on phase,
rt vector
eusen etc
ey detect
signs also
tication'?l,
to detect
| scale of
|a novel
the 64-
and 16-
rm” Liu
histogram
ian visual
tention at
n we saw
ency area
according
> saliency
s such as
earch etc.
ience and
people's
| intensity
computer
lines like
principle
take into
ally have
the visual
ection. In
affic sign
| to detect
safety. In
addition to, a new generated method of saliency map based on
visual contrast is also introduced.
2. METHODOLOGIES
2.4. Two-way Integration Method of Target Detection
Modern physiology and psychology in visual research
shows that the visual process is a integrated process including
both bottom-up and top-down!! !!, so our visual attention model
is two-way integration. In Top-Down phase, we choose Simple
Vector Filter proposed by T. Asakura etc?! as a priori guide.
The filter can highly extract specific colour and remove profile
and have good segmentation results for red, blue and yellow. In
Bottom-Up phase, we use a new saliency analysis method based
on visual contrast.
Visual attention model | Candidate Target
5 = = - L| region area
Top-Down Analyze | | Tl 1
mL | P[RO
| Color-based | =» L| 0 LL—————4 p
||| Segmentation | | | Z | oo [Ti
[ HIVER | | 8 | Analysis DR
/ Input J| t É | it my and (= [T
/ Optical ML = =» —— | Detection of [m
f mes [i Borem-uc 17 | |B] | taskerelated |
= 1908 attention | z TES targets | |
À x | |
[ | Biologically - 2 PE D ue pm
| Yom. J ITE Be
| Mechanism |o alo mem
CS
Fig.1. The detection flowchart of traffic sign
The flowchart of our method is shown in Fig.l. Firstly, the
whole scene will be analyzed by visual attention model based
on two-way integration. We may acquired many saliency areas,
i.e. candidate areas., because there are many object in the scene
and they maybe have the same saliency with task-related objects,
Secondly, these candidate area will be analyzed according to
the shape characteristics of task-related objects to acquire the
needed target area".
2.) Physiology background of saliency analysis
The phenomenon that the retina will strongly respond to large
contrast visual stimulation and the generation mechanism of
visual information in the primary visual cortex can be simulated.
we propose a method generating saliency map according to the
cognitive neuroscience research. The method includes two
layers computational unit and they correspond to simple cell
and complex cell in primate primary visual cortex.
S Unit: Human retina RF will strongly respond to the highest-
contrast visual information, e.g., the center is light but
surrounding is dark. The biological characteristics can be
simulated by using difference operation between central high-
resolution layer and surrounding slow-resolution layer!!!
Primate primary visual cortex contains simple cells and
complex cells. Some studies suggest that the receptive fields of
the simple cell only include a small part of the vision, these
local units must be pooled together by visual system in order to
perceive the target within vision. Complex cells are the
nonlinear spatiotemporal integration of simple cells!'*. In this
research, we use contrast as a local saliency, i.e., the contrast
information given by retinal are local contrast information
generated by simple cell in primary visual cortex. And then,
these local contrast information are integrated together to form
the global information generated by complex cell.
M us |i»
| s Si | 1
‚Max {Si} Y | (ot)
Ime = ——— pt
| |
02) (02 04) (02) (02) (04) (02 (02 0.4
Fig.2. Flowchart of image attention analysis
C Unit: C unit are pooled from the S unit. The computing pool
model is the bridge between complex and simple cells in
primate primary visual cortex. As shown in Fig.2, there are
three basic computing pool model for integrating local units is
proposed in work”. je. Maximum model, Energy model and
Half-wave model. Some experimental evidence in favor of the
max operation has already appeared! !7), So, we choose out the
max model to pool from complex cells to simple cells.
| Bottom-Up Visual attention ~~» Max
| — —» Center-Surround
| i
| S1 Unit C1 Unit
MAX
{Down Sampling) ( Generating / Generating —-—-» Max
«. S1 Unit J cr Unit | | ——» Center-Surround
Fig.3. Flowchart of image attention analysis
Our approach is summarized in Fig.3, within the workflow of image
attention analysis, an input image passes through two parts, S unit, and
C unit. They respectively correspond to simple and complex cells.
After down-sampling and the central-surround operation, we can get
the local contrast map in S unit, and then the max model is employed
to pool from the local contrast map to the global contrast map in C unit.
For further highlighting the saliency areas, the global contrast map is
smoothed with Gaussian filter in order to acquire the saliency map.
Subsequently, the generic threshold segmentation is used to detect the
object in the saliency map, where threshold is three times the average
intensity ofthe saliency map.
2.3 Computational step of saliency map
To sum up, a novel method of detecting the saliency object from the
image has the following specific steps:
1) To generate an image pyramid. Down-sampling the original
image I to create the Gaussian pyramid / (0) , where o is the layer
of the image pyramid. The layer o is set to 4, and thus c e[1..4].
The first layer is the quarter of the size of the original image, the next
layer is a half of the upper layer, and for instance the ration of the
image /(1) and the image /(4) is 1/8.
2) To generate S unit. We use center-surround operation to deal
with 4-layer images in the pyramid and the result is used as the visual
information of S unit. A “center” fine scale ce{1,2,3} and a
“surround” coarser scale 5 € {2,3,4} (s=c+6,6=1) , surround
layer s are interpolated to the scale of the central layer c, and then the
point by point difference operation is used to get three difference