paper1—Machine Learning Approach for Ship Detection using Remotely Sensed Images

1、Tensor Flow
Tensor Flow is a programming system developed by
Google which represents computations as graphs.
Computation Graph is first constructed to train the neural
network and then executed using Sessions. Tensors are arrays
that represent data. Feed and Fetch operations manipulate
data. This library is typical for machine learning purpose
derived from parent Theano.
Tensorflow 是由谷歌以图形表示计算的编程系统。首先构建计算图来训练神经网络然后使用会话执行。张量是代表数据的数组，来操作数据。这个库来自他的父类典型的机器学习库Theano。
A dataset of 8,000 images should be used to train the
model. From the dataset 5,000 will be used for training and
3,000 will be used for testing. Training Dataset consist of
positive images and negative images. Training set is
constructed by tilling the Satellite images into small frames
and then positive image i.e. tile with ship is superimposed
with tile not containing any ship.
一般应使用8000张左右图像来训练模型。拿出5000张数据集用来训练，拿出3000张数据集用来测试。训练数据集包括好的和坏的图像。训练集是通过将卫星图像细分为小帧然后再生成正像即带有船只和不带有船只的。
Neural network uses Weight and biases to train the
dataset. Using back propagation algorithm their values can be
adjusted. The cross-entropy is measuring how inefficient our
predictions are for describing the truth. Tensor Flow
minimizes cross-entropy using the gradient descent algorithm
with a learning rate of 0.5. Gradient descent is a simple
procedure, where Tensor Flow simply shifts each variable a
little bit in the direction that reduces the cost. Iteration will
be performed over the training model to gain accuracy of
minimum 90%. That is out of ten vessels detected one will be
false alarm.
神经网络使用权重和偏置来训练数据集。使用反向传播算法来调整参数值。交叉熵是计算效率和描述我们的真实值。Tensor Flow使用梯度下降算法以0.5的学习速率使交叉熵最小化。梯度下降是一个程序，tensorflow只是将每个变量沿着损失最小的方向移动一小步。将在训练模型上进行迭代以获得至少90％的精确度。就是说检测的十艘船里面有一艘船会有错误警报。
After training, System is given input image to detect the ship
from entire matrix of image. Detected vessel will be enclosed
by rectangle.
训练后，系统会从整个图像矩阵提供输入图像以检测船只。检测到的船只将按矩形被封闭。
Accuracy of System can be calculated by using test images. Softmax Regression technique is used to get probability of tile containing ship in range of 0 and 1[10]. If there is more than 90% probability that given tile has ship, then that tile will be highlighted.
可以通过使用测试图像来计算系统的准确性。 Softmax回归技术用于获取包含船舶的图片在0到1之间的概率。 如果给定的图片有上船的可能性超过90％，则该图片将突出显示。
Another research paper defines image features by their morphological intrinsic characteristics, instead of using their boundary . The method is focused on image structural
information which is collected by applying morphological
operators with a multi-scale approach. Concepts of
morphological profile and derivative of morphological
profile(DMP) are used to create a feature vector from a single
image.
另一篇研究论文通过图像的内在特征来定义图像特征，而不是使用边界。 该方法专注于图像结构通过应用形态学收集的信息
操作员采用多尺度方法。形态特征的概念和形态导数经常被用来从单张图片中创造变量。
The basic idea behind using morphological operators
is to identify even smallest of differences and features from an
image (assuming that the resolution of the image is good
enough).The paper concentrates on linear feature extraction.
One of the methods used for feature extraction is
DAFE(discriminant analysis feature extraction).However it
has a weakness that it is not directly related to the probability
of error in classification. To overcome this, the authors have
proposed another method known as DBFE (decision boundary
feature extraction).The authors showed that DBFE overcomes
many problems and have extended its application to neural
networks.
使用形态运算符的基本思想就是要识别出最小的差异和特征
图片（假设图片的分辨率足够好）。本文着重于线性特征提取。
用于特征提取的方法之一是DAFE（判别分析特征提取）。
缺点是它与分类错误的概率没有直接关系。 为了克服这个问题，作者提出了另一种称为DBFE（决策边界特征提取）的方法。作者表明DBFE克服了许多问题并将其应用扩展到神经网络。
DBFE: The authors showed that discriminant
informative features and discriminant redundant features can
be extracted from the decision boundary itself. The difference
between informative and redundant features defines the
likelihood related to the decision boundary. In discriminant
informative features the likelihood can be defined by a line
normal to the decision boundary whereas in discriminant
redundant features it is defined by a line parallel to the
decision boundary. The likelihood of the two features as
shown in below figure.
DBFE：作者证明了这种区别信息性特征和区别性冗余特征可以
从决策边界本身提取。这种信息性和冗余性的特征的定义可能和决策边界相关。在判别信息特征中，可能性可以由垂直于决策边界的线定义，而在判别冗余特征中，可能性可以由与决策边界平行的线定义。这两个特征的可能性如下图所示。
判别信息性示例。 在此，X表示模式，b表示常数，f（X）表示似然比，t表示决策边界。
DBFE Procedure for Neural Networks (Two-Pattern Class
Case):
Step 1 Train the neural network using all features.
Step 2 For each training sample correctly classified as a class,
find the nearest sample correctly classified as class. Repeat
the same procedure for the samples classified as a class.
Step 3 The line connecting a pair of samples found in Step 2
must pass through the decision boundary, since the pair of
samples are correctly classified differently. By moving along
this line, find the point on the decision boundary or near the
decision boundary within a threshold.
Step 4 At each point found in Step 3, estimate the normal
vector.
DBFE程序提供了两个神经网络示例
步骤1 使用所有的特征训练神经网络
步骤2 对于正确分类为类别的每个训练样本，请找到正确分类为类别的最近的样本。重复相同的程序把样本分类。
步骤3 由于正确地对这对样本进行了不同的分类，因此在步骤2中找到的连接一对样本的线必须穿过决策边界。通过移动这条线，发现决策边界或者接近决策边界的点在一个阈值范围内。
步骤4 对于每个在步骤3中发现的点，估计法线向量。
The rate of false alarms near the ports can thus be minimized
by verifying the output observed from tensor flow using dbfe
as second level classifier.
通过使用DBFE验证从张量流观察到的输出作为二级分类器，可以将端口附近的误报率降至最低。
**

结论

**
Using tensor flow we aim to develop more efficient and
accurate system for detection of ships in real time. This would
therefore increase the security at the coastal areas and improve
the coastal surveillance. Also it would help to detect the lost
ships or any other objects in the ocean. Moreover, if trained
the system accordingly, the moving ships can be detected by
obtaining the videos from a respective satellite. DBFE on the
other hand would give us accuracy in detecting the ships at
ports. The most highlighted disadvantage that has emerged
through research is that the ships at port or near the dock
cannot be correctly classified most of the times. So using
DBFE would therefore solve this problem. Hence this newly
proposed system would be not only efficient but also accurate
enough to detect ships in real time.
我们使用tensorflow的目的是发展对船舶实时监测更有效和精确的系统。这将会增加沿海区域的安全性和提高沿海监测能力。他也能够帮助检测失联的船只和任何在海面的对象。而且，如果相应的训练系统，那么移动的船只能够通过各自的卫星获取的视频进行检测。另一方面，DBFE将使我们能够准确地检测港口的船舶。
通过研究得出的最突出的缺点是，在大多数情况下，无法正确分类港口或码头附近的船只。因此使用DBFE将解决这个问题。因此这个新的提议在实时船舶检测中不仅高效而且精确度足够好。

如有侵权，请联系删除。