2 Recent Advances in Electrical & Electronic Engineering, 2019, Vol. 12, No. 1 Tao et al.
(a) (b) (c)
d) (e)
Fig. (1). Illustration of segmentation of the target region and shadow. (a) Original image (b) Equalized image (c) Results after mean filter (d)
Preliminary extraction results. (e) Extracted region and shadow.
Therefore, JSR can better eliminate the discrimination
observed in the target region and shadow to improve the
ATR performance. Finally, the target type of the test image
is decided, based on the total reconstruction errors from JSR.
The remainder of this paper is organized as follows. Sec-
tion 2 introduces the segmentation of the target region and
shadow and the feature construction based on them. In Sec-
tion 3, the detailed implementation of the proposed target
recognition method is described. Extensive experiments are
conducted on the Moving and Stationary Target Acquisition
and Recognition (MSTAR) dataset under different operating
conditions in Section 4 to evaluate the effectiveness and ro-
bustness of the proposed method. Finally, in Section 5, con-
clusions are drawn based on the experimental results.
2. MATERIALS AND METHODS
2.1. Feature Construction Based On Target Region And
Shadow
2.1.1. Extraction of the Target Region and Shadow
The target region describes the backscattering character-
istics of the target whereas the shadow reflects the occlusion
between the target and ground at a special radar view angle.
Therefore, the target region generally contains pixels with
large intensities. In contrast, the shadow is always filled with
much lower pixel values than the background and target re-
gion [6]. In this study, the target region and shadow are ex-
tracted by a simple but effective thresholding algorithm. The
detailed implementation is illustrated in following steps:
Step 1: Equalize the intensities of the original image into
the range of [0, 1];
Step 2: Perform the mean filter on the equalized image
with the kernel size of 3 × 3;
Step 3: Set T1=0.85 and T2=0.15; the pixels with higher
intensities than T1 are assigned to the target region and those
with intensities between T2 and T1 are shadow pixels;
Step 4: The morphological opening operation is used to
remove the false alarms and the morphological closing oper-
ation is used to connect the target region and shadow.
An illustration of the extraction of the target region and
shadow are given in Fig. (1), where Fig. (1a) shows the orig-
inal image with the size of 128 × 128 pixels from the
MSTAR dataset. The equalized image is shown in Fig. (1b).
Fig. (1c) shows the results after the mean filter. The thresh-
olding method produces the preliminary segmentation results
as Fig. (1d). Finally, the smoothed target region and shadow
after the morphological operations are presented in Fig. (1e).
As shown, both the target region and shadow provide physi-
cally relevant descriptions of the target, e.g., physical sizes
and geometrical shape. In addition, the relative positions of
the target region and shadow reflect the radar view angles
and target poses, which are also discriminative for target
recognition.
2.2. Feature Extraction
EFDs have been widely used to describe the complicated
target contour. Based on the extracted target region and
shadow, the standard Sobel operator [38] is performed to
achieve their outlines. For a smoothed and closed outline
curve