IEEE Geoscience and Remote Sensing Magazine - December 2021 - 11

in labeling data and is available to quickly create largescale
object-detection data sets for specific applications.
The precise pixelwise segmentation mask is an annotation
method, wherein each pixel in the image is assigned a category
label, such as forest, farmland, road, or background.
This type can be applied in scenarios in which the environmental
context is important. This type of annotation
requires more expert knowledge and labor to be successful.
Due to the massive number of object categories and
instances, complex backgrounds, and a large data volume
of HRRS images, the precise pixelwise segmentation mask
annotation is rarely used in large-scale remote sensing target
detection.
There are two types of widely used BB annotation methods:
horizontal BBs (HBBs) and orientation (rotation) BBs
(OBBs). In Figure 1, HBBs (the axis-aligned rectangle) were
first used to localize objects. However, objects in HRRS images
often appear in arbitrary orientations and may be densely
distributed. In some extreme but common scenarios, this
annotation method involves both the background and targets
of interest; it cannot accurately or compactly outline the
locations of objects and may decrease detector performance.
The annotation method of the OBBs, which can be regraded
to add angles to the HBBs, is utilized to gain a tight bounding
for the rotation object. For this article, we review mainly
methods with these two types of BB annotations.
DIFFICULTIES AND CHALLENGES IN SMALL,
WEAK OBJECT DETECTION
Relevant works for small, weak object detection of infrared
images started to appear long ago, when the spatial resolution
of remote sensing images was relatively low and infrared
images were the main data source for object detection.
Many works have addressed solutions to such problems
[28]-[32]. Related works covering the analysis of object
detection in infrared images [33]-[35] originally defined a
small object as one with a total spatial extent of fewer than
80 pixels (a width of fewer than nine pixels), which is less
than 0.2% in an image of 256 × 256 pixels. As shown in
Figure 2, the long distance of imaging means that the target
takes up only a few dozen pixels in the imaging plane,
presenting as a small target. Objects of this kind are basically
shapeless and have no available texture features. Small
objects are usually characterized as having a low SNR, small
size, and no adequate structure information for the undulant
clutter and imaging distance. These characteristics make
small targets very difficult to detect, and small targets are
easily overwhelmed by a bright background [6]. Therefore, a
small, weak object is more formally defined as 1) small: the
scale of the target is small, or the target's proportion of the
total images is low; and 2) weak: the features of the target are
insufficient and easily affected by its background.
Thanks to the acquisition of HRRS images and the requirements
of the related applications, small, weak object
detection has attracted increasing attention. Although numerous
efforts have been made to develop detectors and
DECEMBER 2021 IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE
benchmark data sets for promoting the development of
small, weak object detection in HRRS images, there is still
no consensus on the definition of small, weak object detection.
Kang et al. [16] proposed a complex background
benchmark wherein vehicles were set as the small objects.
VEDAI (for " vehicle detection in aerial imagery " ) [17] is a
data set created for small target detection, but the authors
did not propose a specific definition for a small object. Xia
et al. [18] proposed a new large-scale benchmark for HRRS
image detection, dividing the object instances into three
classes according to the pixel width of their BBs: small for
a width from 10 to 50, middle for a width from 50 to 300,
and large for a width of greater than 300. This was the first
work to clearly define a scale for small objects in HRRS
images. For the aspect of the weak feature response of objects,
no related work has proposed sufficient discussion or
drawn a clear conclusion.
In this section, we comprehensively consider the factors
that affect detection performance and then summarize the
difficulties of and challenges to small, weak object detection
in HRRS images. In Table 1, each influencing factor is
examined from the three aspects of image quality, object
variations, and complex context as follows.
1) Image quality: In the process of HRRS image acquisition, the
imaging environment, satellite platform, optical system,
FIGURE 2. Small, weak object detection in infrared images.
TABLE 1. THE CHALLENGES AFFECTING SMALL,
WEAK OBJECT DETECTION.
THREE
ASPECTS
Object
variations
SPECIFIC CONTENT
Image quality Mixed noise, patch missing, occlusion caused by
cloud, fuzzy, shadow, and multisource data
Small size, high intraclass variations, a change of the
object features caused by illumination and background,
antagonism of the background and the target,
a lack of annotation samples, nonuniform distribution,
and an imbalance of positive and negative
training examples
Complex
context
Many types and quantities of background targets and
complex distribution patterns
11
IEEE Geoscience and Remote Sensing Magazine - December 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - December 2021
