基于最小方向对比度的300 FPS显著目标检测

需积分: 9 151 浏览量更新于2024-08-26 收藏 2.9MB PDF 举报

"这篇论文提出了一种新的显著性对象检测算法——基于最小方向对比度（Minimum Directional Contrast, MDC）的方法，适用于300帧每秒的高速处理。该方法改进了传统的全局对比度计算方式，考虑了对比度的空间分布，提升了显著性检测的准确性与效率。" 在计算机视觉领域，显著性目标检测是一项关键任务，其目标是识别并定位图像中最具吸引力或最突出的区域，这些区域通常包括图像的主要焦点或兴趣点。传统的显著性检测算法常常基于全局对比度，即计算目标区域或像素与其周围环境在颜色、亮度等方面的差异。然而，这种单纯基于全局对比度的方法往往忽视了对比度在空间上的分布特征。黄晓明和张宇瑾两位作者提出的“300-FPS Salient Object Detection via Minimum Directional Contrast”论文，针对这一问题进行了创新。他们观察到前景像素通常在各个方向上都有高对比度，因为它们被背景包围；而背景像素往往至少在一个方向上的对比度较低，因为它们需要与背景相连。基于这个观察，他们提出了一个新的原始显著性度量——最小方向对比度（MDC）。MDC通过计算每个像素从不同方向的对比度，有效地捕捉了像素的显著性信息。为了实现高效的计算，论文中还提出了使用积分图像进行O(1)复杂度的MDC计算方法。这种方法使得对于QVGA分辨率的输入图像，仅需1.5毫秒就能完成处理，极大地提高了处理速度，满足了实时处理的需求，如在300帧每秒的视频流中进行显著性检测。在显著性后处理阶段，MDC结果可以与其他特征结合，如边缘信息、纹理分析等，进一步提升检测的准确性和鲁棒性。此外，这种方法可能对实时监控、自动驾驶、人机交互等应用具有重要意义，因为它能够在不影响性能的情况下，快速且准确地提取图像中的显著目标。这篇论文通过引入最小方向对比度的概念，为显著性目标检测提供了新的视角和高效的计算方法，有助于推动该领域的技术发展。

HUANG AND ZHANG: 300-frames/s SALIENT OBJECT DETECTION VIA MDC 4245

contrast also named as uniqueness is deﬁned as Euclidean

color distance to entire image with Gaussian spatial weight:



j=1

||c

− c

= c



j=1

−2c



j=1



j=1

(2)

where c

is the average color of the region or pixel r

.Since

is a Gaussian spatial weight function, using Gaussian

blurring kernel on c

and c

, the computational complexity for

each pixel can be reduced from O(M) to O(1). Nevertheless,

Gaussian blurring still needs high time expenditure.

In above global contrast based work, saliency is measured

by the sum of contrast from the entire image, some background

regions also show high saliency due to large contrast contribu-

tion from the object. An example is shown in Fig. 1. (e)(f)(g),

the background pixel at yellow crossing and the foreground

pixel at red crossing demonstrate comparable saliency. In this

paper, we consider spatial distribution of contrast, and adopt

MDC as saliency metric, contrast contribution from object to

background can be suppressed.

Besides, RC [18] adopt graph-based segmentation [34]

which needs 60ms for each image, expensive time cost for

region segmentation limits speed performance.

The boundary and connectivity priors [21] are also shown

to be effective in salient object detection. These priors

assume that background regions are usually connected to the

image boundary. Geodesic distance [21] and Minimum Barrier

Distance (MBD) [46] are widely used distance transform to

measure regions connectivity to the image boundary. In [40],

authors propose one approximate MBD implementation with

raster scan. Three passes scan on each color channel also limit

the speed performance. Another approximation method on

minimum spanning tree (MST) in presented in [45], additional

time cost to build tree leads to worse speed performance

than [40]. Both these two algorithms show that MBD is more

robust to noise and blur than the geodesic distance.

Recently, deep learning has achieved great successes

in many computer vision tasks. Some researchers have

already applied deep neural networks to saliency detection.

Wang et al. [47] use a CNN to predict saliency for each pixel

in local context, then reﬁne the saliency on object proposal

over the global view. Zhao et al. [48] consider global and

local context simultaneously in a multi-context CNN, then

combine them to predict saliency. Deep learning based works

both achieve better performance with very low speed.

III. S

ALIENT OBJECT DETECTION METHOD

In this section, we present an efﬁcient salient object detec-

tion method. We ﬁrst propose one raw saliency metric MDC

which considers spatial distribution of contrast. Next, an O(1)

implementation of MDC is proposed. In post processing,

saliency smoothing and enhancement will be introduced.

A. Minimum Directional Contrast (MDC)

To measure the saliency of a region or pixel, contrast is the

most frequently used feature. Global contrast is widely studied

in salient object detection which considers the color difference

between the target region or pixel and the entire image.

Global contrast can be calculated at pixel or region level.

Region level method needs higher time cost for image seg-

mentation. In RC [18], graph-based segmentation [34] needs

about 60ms for each image on MSRA-10K dataset [18] [33].

In SLIC superpixel segmentation based method SF [20],

segmentation needs about 110ms. In order to achieve higher

speed performance, we adopt pixel level saliency detection

which needs not time cost for region segmentation.

As discussed in introduction part, in previous global contrast

based methods, saliency is simply measured by the sum of

contrast from the entire image [18], [20], or deﬁned as contrast

with the average image color [17]. Spatial distribution of

contrast is neglected.

In this paper, we further detail analyze contrast from differ-

ent spatial directions. If the target pixel i is regarded as the

center of view, the entire image can be divided into several

regions based on their location w.r.t. pixel i,i.e.,topleft(TL),

top right (TR), bottom left (BL), bottom right (BR). Directional

contrast (DC) from each region  can be calculated as:

i,









j∈



ch=1

i,ch

− I

j,ch

)

(3)

where I denotes one input image with K color channels in

CIE-Lab color space. Fig. 3 (a) is one input image, two target

pixels are shown in Fig. 3 (b), one foreground pixel at the

red crossing in the top row, and one background pixel at the

yellow crossing in the bottom row. The entire image is simply

divided into four regions by red or yellow line. DC result of

two target pixels is shown in Fig. 3 (c).

From DC result in Fig. 3 (c), we can ﬁnd the distribution

of DC differ greatly between foreground and background

pixel. In the top row, foreground pixel shows high DC in

almost all directions, minimum directional contrast (MDC)

still has high value. In the bottom row, background pixel

demonstrates very low DC in direction bottom right and high

DC in other directions, MDC will show very low value. More

generally, since foreground pixel is usually surrounded by the

background, it often has high contrast from all directions,

MDC shows high value. On the contrary, the MDC of a

background pixel is usually small, as it has to connect to the

background through one of the directions. This suggests that

we can deﬁne MDC which means the minimum contrast from

all directions as the raw saliency metric:

S(i) = min

 ∈ TL,TR,BL,BR

i,







min

 ∈ TL,TR,BL,BR

(



j∈



ch=1

i,ch

− I

j,ch

)

(4)

The MDC of two target pixels are shown in Fig. 3 (c),

the foreground pixel at red crossing shows obviously higher

MDC than the background pixel at yellow crossing. The MDC

based raw saliency of all pixels are shown in Fig. 3 (d).

In previous global contrast based methods (HC [18],

RC [18], SF [20]), saliency is simply deﬁned as the sum of

剩余11页未读，继续阅读

myhasdfg

粉丝: 1
资源: 6

基于最小方向对比度的300 FPS显著目标检测

20200894_RGB-D Salient Object Detection A Survey.pdf

SALIENT OBJECT DETECTION VIA BACKGROUND CONTRAST

Background Prior-based Salient Object Detection via Adaptive Figure-Ground Classification

SALIENT OBJECT DETECTION VIA OBJECTNESS MEASURE.pdf

Depth-aware salient object detection using anisotropic center-surround difference

Optimizing the F-measure for Threshold-free Salient Object Detection (ICCV 2019)

层次分析matlab代码-Co-saliency-Detection-via-Co-salient-Object-Discovery-and-

Frequency-tuned salient Region Detection.zip_gonegfi_matlab_regi

Frequency-tuned Salient Region Detection

Salient-Object-Detection_CVPR2014.zip_图形图像处理_matlab_

最新资源