4
Q .W a n ge ta l .
2.2 Feature Fusion and the Dimension Reduction
Existing deep learning pre-trained CNN models are used as the feature extractor to extract the feature from the final FC layers (include the last classification
layer). Although the feature can be utilized to train the classifier directly, effective features extracted by only a single CNN pre-trained model are not enough,
which will lead to a disappointing effect. Hence, an efficacious way to solve this
question, feature fusion, is proposed.
Feature fusion: Deep feature fusion is a new solution to handle complex
data. In 2014, two MIT engineers developed deep feature synthesis [10]. Most
prediction decisions rely on the features descriptors based on input images in the
vision classification tasks. Hence, it is necessary to overcome the obstacles of data
dependence. Feature fusion refers to concatenating global features descriptors
extracted from several different pre-trained CNN models. Vectors obtained by
these models expand the dimension of the final vector by vector-spliced which is
an efficient method to mitigate data dependence. In addition, the key advantage
of feature fusion is new features obtained in this process, which can improve the
performance of classification.
Reduce dimension: Dimensionality reduction, as the name implies, means
feature selection and feature extraction. Since principal component analysis [11]
was successfully proposed, applying PCA transformation to reduce dimension
becomes a mainstream tendency. In the proposed method, PCA, as an important
processing technology, is introduced for feature descriptors distinguishing and
dimension reducing. With PCA more consummate, the field of data it can handle
becomes wider. It is also the main technology for compressing the time of training
process without losing the quality of a model. The main process for PCA is to
transform the original data onto a set of linearly independent representations of
each dimension through linear transformation, which reduces the dimension of
input data set while maintaining the feature of the largest contribution of the
data set in the data set. The final result is the key feature components of all the
features.
3 Experiments and Analyses
3.1 UC Merced Land Use Dataset Description
In this section, we investigate the performance of the proposed methods on the
“UC Merced Land Use” dataset
1 [12] extracted from large images from the
US Geological Survey National Map Urban Area Imagery collection for various
urban areas around the country. This dataset contains 21 classes. Each class
includes 100 images with the size of 256 × 256 pixels in the color space of red–
green–blue with different space structure, color distribute, region cover, and
object cover. Every image is operated by rotating.
1 http://vision.Ucmerced.edu/datasets/landuse.html.
Q .W a n ge ta l .
2.2 Feature Fusion and the Dimension Reduction
Existing deep learning pre-trained CNN models are used as the feature extractor to extract the feature from the final FC layers (include the last classification
layer). Although the feature can be utilized to train the classifier directly, effective features extracted by only a single CNN pre-trained model are not enough,
which will lead to a disappointing effect. Hence, an efficacious way to solve this
question, feature fusion, is proposed.
Feature fusion: Deep feature fusion is a new solution to handle complex
data. In 2014, two MIT engineers developed deep feature synthesis [10]. Most
prediction decisions rely on the features descriptors based on input images in the
vision classification tasks. Hence, it is necessary to overcome the obstacles of data
dependence. Feature fusion refers to concatenating global features descriptors
extracted from several different pre-trained CNN models. Vectors obtained by
these models expand the dimension of the final vector by vector-spliced which is
an efficient method to mitigate data dependence. In addition, the key advantage
of feature fusion is new features obtained in this process, which can improve the
performance of classification.
Reduce dimension: Dimensionality reduction, as the name implies, means
feature selection and feature extraction. Since principal component analysis [11]
was successfully proposed, applying PCA transformation to reduce dimension
becomes a mainstream tendency. In the proposed method, PCA, as an important
processing technology, is introduced for feature descriptors distinguishing and
dimension reducing. With PCA more consummate, the field of data it can handle
becomes wider. It is also the main technology for compressing the time of training
process without losing the quality of a model. The main process for PCA is to
transform the original data onto a set of linearly independent representations of
each dimension through linear transformation, which reduces the dimension of
input data set while maintaining the feature of the largest contribution of the
data set in the data set. The final result is the key feature components of all the
features.
3 Experiments and Analyses
3.1 UC Merced Land Use Dataset Description
In this section, we investigate the performance of the proposed methods on the
“UC Merced Land Use” dataset
1 [12] extracted from large images from the
US Geological Survey National Map Urban Area Imagery collection for various
urban areas around the country. This dataset contains 21 classes. Each class
includes 100 images with the size of 256 × 256 pixels in the color space of red–
green–blue with different space structure, color distribute, region cover, and
object cover. Every image is operated by rotating.
1 http://vision.Ucmerced.edu/datasets/landuse.html.