Aerospace and Electronic Systems Magazine July 2018 - 40

Hybrid Guidance System Based on Infrared and Millimeter Waves

Figure 8.

Above: Traditional extraction of the image feature. (a) Original diagram. (b) Scaling down. (c) Feature
extraction. Below: Fast extraction of the image feature. (d) Original diagram. (e) Feature extraction. (f)
Scaling down.

Although artificial feature selection provides a high-dimensional floating-point feature, feature extraction related to storage
characteristics and the feature matching process is a time-consuming task. The hashing methods, of which the most commonly used
are the hierarchical method and the Hashi code method [21], convert feature characteristics into binary characteristics and solve the
storage problem with the required accuracy and at a reasonable
cost. Extraction results for target features are shown in Figure 8.
In this article, a classification and recognition method for timesensitive targets based on depth CNN, combined with feature coding and feature selection, is proposed. In the proposed method,
when the CNN feature is extracted in the detection phase, the feature is reused directly in the recognition phase. The specific steps
of CNN feature extraction are presented in Figure 9.
First, the detected image is fed into the input of the deep CNN
model to extract the features, and the underlying features with a
strong expressive ability are obtained as a CNN output. Second, the

extracted feature is used to increase
the ability of feature description.
Finally, the target feature selection
technique is used to reduce the dimensionality of a high-dimensional
feature to remove redundancy. Finally, features are sent to the classifier.
With the aim of acquiring
enough live images, Internet resources were used to find keywords
contained in images of interest.
Then, a weakly supervised learning
technology was employed to determine target position, label box, and
label categories. Afterward, assuming that a sample is only part of the
scratch, the target position and the
label of unlabeled images were determined using the weak supervised
learning technology.

MILLIMETER WAVE AUTOMATIC TARGET RECOGNITION
BASED ON A MORPHOLOGICAL NEURAL NETWORK
The morphological neural network (MNN) is a nonlinear neural
network that combines mathematical morphology and the classical
neural network approach. Our main idea was to use a morphological multiplication operator and an addition operator instead of a
classical neural network.
Compared to classical neural networks, MNNs have the following advantages:
1. Ability to generate hidden layers automatically in the middle of
the network without previous definition and avoid the classical
neural network parameters
2. Ability to converge in a few steps and achieve perfect classification of training samples

Figure 9.

Classification and recognition method for time-sensitive targets based on depth CNN, combined with feature coding and feature selection.

IEEE A&E SYSTEMS MAGAZINE

JULY 2018

Table of Contents for the Digital Edition of Aerospace and Electronic Systems Magazine July 2018