processing. Radar echo signal is a one-dimensional discrete sequence. A radar data file
usually contains data of several distance units for a certain observation time, which
determines that there are many input forms under the conditions of translation
invariance and locality required for classification using convolutional neural network.
In order to better use the deep learning method for target detection in radar sea clutter,
several feasible input forms are introduced in this paper.
2 Pretreatment Mode
2.1 One-Dimensional Form
Common radar data sets, such as IPIX data sets [5] and CSRI data sets [6], usually store
the values of sampling points in matrix form. In the longitudinal direction, the sampling
points are determined by the pulse repetition frequency (PRF) and the observation time,
while in the transverse direction, they correspond to each distance unit. Each column of
data is a discrete one-dimensional time series, so it can input in the form of onedimensional time series. In addition, there are many commonly used processing
methods for time series, such as time-frequency transform, joint time-frequency
analysis, etc. [7]. These methods can be used to preprocess the data before the input
network, so as to improve the discrimination between target data and clutter data, and
then improve the accuracy.
2.1.1 Direct Input
Each range unit of radar data file usually contains thousands of sampling points (such
as CSIR data set) or tens of thousands of sampling points (such as IPIX data set). Due
to the limited amount of data, it is unrealistic to input all sampling points of a distance
unit directly. A certain echo data can be intercepted on a distance unit with a fixed
preset length as a data sample. As shown in Fig. 1, each distance unit is slipped for
interception, and after interception, it is transferred to the next distance unit for
interception until enough data samples are intercepted. When the amount of data is
insufficient, the sliding step can be smaller than the preset length; when the amount of
data is sufficient, the sliding step can be larger than the preset length. Compared with
the target detection in clutter by input radar P-display images, direct input of radar echo
sequence samples can make better use of the original information contained in the echo
sequence, but requires a higher signal-to-clutter ratio as support. At the same time,
according to the specific data set, we need to determine the value of the preset length
through a large number of experiments, which is too short to contain enough data for
classification, and too long to ensure that there are enough samples to form a training
set to achieve convergence. So it is necessary to find a balance between the amount of
information in each data sample and the amount of data in the data set. In addition, in
the process of sample interception, it is inevitable that the texture information of radar
echo data will be destroyed at some interceptions.
Pretreatment Method of Deep Learning Method for Target …
609
usually contains data of several distance units for a certain observation time, which
determines that there are many input forms under the conditions of translation
invariance and locality required for classification using convolutional neural network.
In order to better use the deep learning method for target detection in radar sea clutter,
several feasible input forms are introduced in this paper.
2 Pretreatment Mode
2.1 One-Dimensional Form
Common radar data sets, such as IPIX data sets [5] and CSRI data sets [6], usually store
the values of sampling points in matrix form. In the longitudinal direction, the sampling
points are determined by the pulse repetition frequency (PRF) and the observation time,
while in the transverse direction, they correspond to each distance unit. Each column of
data is a discrete one-dimensional time series, so it can input in the form of onedimensional time series. In addition, there are many commonly used processing
methods for time series, such as time-frequency transform, joint time-frequency
analysis, etc. [7]. These methods can be used to preprocess the data before the input
network, so as to improve the discrimination between target data and clutter data, and
then improve the accuracy.
2.1.1 Direct Input
Each range unit of radar data file usually contains thousands of sampling points (such
as CSIR data set) or tens of thousands of sampling points (such as IPIX data set). Due
to the limited amount of data, it is unrealistic to input all sampling points of a distance
unit directly. A certain echo data can be intercepted on a distance unit with a fixed
preset length as a data sample. As shown in Fig. 1, each distance unit is slipped for
interception, and after interception, it is transferred to the next distance unit for
interception until enough data samples are intercepted. When the amount of data is
insufficient, the sliding step can be smaller than the preset length; when the amount of
data is sufficient, the sliding step can be larger than the preset length. Compared with
the target detection in clutter by input radar P-display images, direct input of radar echo
sequence samples can make better use of the original information contained in the echo
sequence, but requires a higher signal-to-clutter ratio as support. At the same time,
according to the specific data set, we need to determine the value of the preset length
through a large number of experiments, which is too short to contain enough data for
classification, and too long to ensure that there are enough samples to form a training
set to achieve convergence. So it is necessary to find a balance between the amount of
information in each data sample and the amount of data in the data set. In addition, in
the process of sample interception, it is inevitable that the texture information of radar
echo data will be destroyed at some interceptions.
Pretreatment Method of Deep Learning Method for Target …
609