This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 50, NO. 9, SEPTEMBER 2015 1
A 30 µW 30 fps 110 × 110 Pixels Vision Sensor
Embedding Local Binary Patterns
Andrew Berkovich, Student Member, IEEE, Michela Lecca, Leonardo Gasparini, Member, IEEE,
Pamela A. Abshire, Member, IEEE, and Massimo Gottardi, Member, IEEE
Abstract—We present a 110 × 110 pixel vision sensor that com-
putes the Local Binary Patterns (LBPs) of an imaged scene with a
power consumption of 30 µW at 30 fps. The LBP of a given pixel is
a binary vector, encoding the direction and sign of image contrast
with respect to its neighbors. Each LBP provides a visual descrip-
tion of an image's local structure that is widely used for texture
and object recognition. In the sensor proposed here, each pixel de-
tects its corresponding LBP with respect to its four neighboring
pixels and saves this information into a digital map using 6 bits to
encode each pixel. The operation is executed during the exposure
time and requires 83 pW/pixel frame to be computed. The chip
is implemented in a 0.35 µm CMOS featuring 34 T square pixels
with 26 µm pitch. We illustrate some examples of image descrip-
tion based on the LBPs output by the sensor.
Index Terms—Active pixel sensors, image sensors, local binary
pattern, low-power vision, visual processing.
I. INTRODUCTION
U
LTRA-LOW power imaging has become a challenging
topic for many applications such as mobile devices,
wireless sensor networks and wearable electronics. While solu-
tions for limiting the power consumption of singular electronic
components have been adopted, bringing vision technology to
ultra-low power consumption is still an open issue [1], [2]. In
fact, vision systems often require large data acquisition and
massive parallel processing, which hardly fit with low power
consumption. Minimizing the power consumption of individual
components does not sufficiently reduce the power budget of
such a system. A more effective solution adopts an integrated
approach—one that merges the image acquisition and image
processing tasks which are usually separated in standard vision
systems. Hardware and software integration is particularly
convenient in the case of parallel image processing, i.e. when
the system needs to repeat the same operations over each
pixel. Embedding this visual processing in hardware, instead
Manuscript received February 04, 2015; revised May 06, 2015; accepted June
05, 2015. This paper was approved by Associate Editor Gyu-Hyeong Cho. The
work was supported in part by the Project EnerViS, “Energy Autonomous Low-
Power Vision Systems,” within the Provincia Autonoma di Trento and Univer-
sity of Maryland R&D Cooperation Program 2012.
M. Lecca, L. Gasparini, and M. Gottardi are with Fondazione Bruno Kessler,
Povo (TN), Italy (e-mail: lecca@fbk.eu; gasparini@fbk.eu; gottardi@fbk.eu).
A. Berkovich and P. A. Abshire are with the Department of Electrical and
Computer Engineering, Institute for Systems Research, University of Maryland,
College Park, MD 20740 USA (e-mail: asb77@umd.edu; pabshire@umd.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JSSC.2015.2444875
Fig. 1. An example of LBP with radius and . The picture reports
the gray values of the central pixel (in red) and of its 4 neighbors (in black).
The neighbors have been sampled in clockwise order: the number (from 0 to
3) close to each pixel indicates the sampling order. The corresponding LBP is
and its rotation invariant version is . See text and Eq. (1) for more
details.
of software, leads to a more efficient, custom system with
reduced power consumption. Using this approach, signals are
often filtered and/or binarized either at the pixel-level [3], [4]
or array-level [5] and thus have the additional benefit of not
needing high performance A/D converters. The main drawback
is that a custom sensor heavily limits system flexibility, i.e.
the information delivered by the sensor cannot be used for
any visual task. In general, a good trade-off between energy
efficiency and usability must be defined taking into account the
application scenarios.
In this work, we describe the architecture of a low power
vision sensor that embeds the computation of the local binary
patterns (LBPs) of each pixel in a cluster of neighboring pixels
(Fig. 1). The LBPs are visual features measuring a binary, direc-
tional, spatial contrast in a neighborhood of each pixel. They de-
scribe image micro-structures, such as edges, corners, lines, or
flat regions, and are invariant against changes in light intensity.
Usually, they are normalized to be insensitive to in-plane rota-
tions. Originally introduced to describe image textures [6], the
LBPs are widely applied to many computer vision tasks, such as
face analysis and detection [7], [8], fingerprints recognition [9],
video background subtraction [10], and image retrieval [11]. For
these applications, embedding the LBPs' computation in hard-
ware reduces the computational load of the processor and the
energy consumption of the entire system.
Our vision sensor captures images with a 110 110 pixel
array and has a power consumption of 30 W at 30 fps. Each
pixel detects its corresponding LBP with respect to its four
neighboring pixels and stores this information into a digital map
using 6 bits to encode each pixel. The operation is executed
during the exposure time and requires 83 pW/pixel frame
0018-9200 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.