2
Q .W a n ge ta l .
descriptors, premeditating the feature extraction as a critical step to obtain a
great classification performance. Existing RSI features extraction research methods are primarily divided into local features extracting methods and global features extracting methods. For the former, there exist two types of methods,
corner feature extraction methods and edge feature extraction methods. In the
past, corner feature descriptor was the main job in feature extraction, such as
features from accelerated segment test (FAST) [2], oriented fast and rotated
brief (ORB) [2]. However, the curves of edge are discontinuous. Hence, they
have limitations on the description of scene semantic information.
Numerous algorithms have been introduced for edge feature descriptor
extraction. Signature of Histograms of Orientations (SHOT) [3] and Raster Operations Units (ROPS) [4] are based on the histogram statistics for local descriptors
extraction, but they are low expression for semantic information in most of the
experiments. In summarize, local descriptors extracting methods are considerably depending on the experiences of researchers, and higher requirements for
extracted feature descriptors are necessary for these methods.
Among all kinds of methods to extract features, comprehensive feature
descriptors can be obtained by deep learning-based methods, which help to
enhance the ability of image description greatly. In 2012, Krizhevsky et al. [5]
proposed a sensational deep learning neural network for image classification,
which picks up the 2012 image recognition contest champion. From then on, the
CNN architecture detonated the application boom of neural networks. Simultaneously, more deep CNN architectures were proposed after AlexNet [5], such
as VGGNet [6], ResNet [7]. Due to the convenience of extracting features and
the better result of classification, they are applied widely in many aspects of
image recognition. Liu et al. [8] proposed to concatenate features extracted from
convolutional layers of CaffeNet and VGG-VD16 to deep descriptors.
In this paper, we propose a method to accelerate the speed of RSI classification model training with significant performance increase. Specifically, we first
obtain features from the last FC layer of two deep CNN models. Then we reduce
the feature vector dimension utilizing PCA transformation. For accelerating the
training speed and exploring a better effect on classification result, we propose
an optimize and simple way to generate feature vector by concatenating the
features from two types of different CNN models.
2 Proposed Method
The details of the proposed method are shown in Fig. 1. The main content consists of the following three parts: global descriptors extracting, feature fusion,
and reducing dimensions. The procedure is organized as follows. The first part
extracts global feature descriptors from the last fully connected layers, including
a sample of the training set and the used pre-train CNN. Two types of pretrain CNN, VGGNet-16 and ResNet-50, are introduced to extract global feature
descriptors. The second part is the details of the proposed method.
Q .W a n ge ta l .
descriptors, premeditating the feature extraction as a critical step to obtain a
great classification performance. Existing RSI features extraction research methods are primarily divided into local features extracting methods and global features extracting methods. For the former, there exist two types of methods,
corner feature extraction methods and edge feature extraction methods. In the
past, corner feature descriptor was the main job in feature extraction, such as
features from accelerated segment test (FAST) [2], oriented fast and rotated
brief (ORB) [2]. However, the curves of edge are discontinuous. Hence, they
have limitations on the description of scene semantic information.
Numerous algorithms have been introduced for edge feature descriptor
extraction. Signature of Histograms of Orientations (SHOT) [3] and Raster Operations Units (ROPS) [4] are based on the histogram statistics for local descriptors
extraction, but they are low expression for semantic information in most of the
experiments. In summarize, local descriptors extracting methods are considerably depending on the experiences of researchers, and higher requirements for
extracted feature descriptors are necessary for these methods.
Among all kinds of methods to extract features, comprehensive feature
descriptors can be obtained by deep learning-based methods, which help to
enhance the ability of image description greatly. In 2012, Krizhevsky et al. [5]
proposed a sensational deep learning neural network for image classification,
which picks up the 2012 image recognition contest champion. From then on, the
CNN architecture detonated the application boom of neural networks. Simultaneously, more deep CNN architectures were proposed after AlexNet [5], such
as VGGNet [6], ResNet [7]. Due to the convenience of extracting features and
the better result of classification, they are applied widely in many aspects of
image recognition. Liu et al. [8] proposed to concatenate features extracted from
convolutional layers of CaffeNet and VGG-VD16 to deep descriptors.
In this paper, we propose a method to accelerate the speed of RSI classification model training with significant performance increase. Specifically, we first
obtain features from the last FC layer of two deep CNN models. Then we reduce
the feature vector dimension utilizing PCA transformation. For accelerating the
training speed and exploring a better effect on classification result, we propose
an optimize and simple way to generate feature vector by concatenating the
features from two types of different CNN models.
2 Proposed Method
The details of the proposed method are shown in Fig. 1. The main content consists of the following three parts: global descriptors extracting, feature fusion,
and reducing dimensions. The procedure is organized as follows. The first part
extracts global feature descriptors from the last fully connected layers, including
a sample of the training set and the used pre-train CNN. Two types of pretrain CNN, VGGNet-16 and ResNet-50, are introduced to extract global feature
descriptors. The second part is the details of the proposed method.