1.2-dimensional data
from the frame convolutional neural network, how to use it for image recognition, object detection, and face recognition neural network converts
the image data even when much of the discussion, the 2D data of one sense are considered the image is so common to
many of the hands of thought is not only limited to 2D images, and even can be extended to 1D, and even 3D data
2D convolution on, the input image may be a 14 × 14, and uses a 5 × 5 filter convolution
followed seen how 14 × 14 image is convolved with a filter of 5 × 5
by this operation will be output in the 10 × 10
If multiple channels, such as 14 × 14 × 3, then the matched filter may be a 5 × 5 × 3
If multiple filters, such as 16, you finally get is 10 × 10 × 16
2.1-dimensional data
proved to be too early idea for the one-dimensional data
For example, an EKG signal is left, or an electrocardiogram (Electrocardiogram)
when one electrode is placed on the chest, the measurement electrodes through the weak current caused by the heartbeat chest
because of heart beat, a weak radio wave can be generated set of measurement electrodes
this is the person EKG heartbeat generation, each peak corresponds with a heartbeat
To use the EKG signal, such as medical diagnostics
that 1-dimensional data processing, because the data is EKG instantaneous voltage of each time sequence is composed of
the size of the input is not a 14 × 14, 14 may be in a size input
In such a situation may require the use of a one-dimensional convolving filter
requires only a 1 × 5 filter instead of a 5 × 5
When the use of the one-dimensional convolution of the signal, to find a 14-dimensional data to 5-dimensional convoluted data, and generates a 10-dimensional output
Again If multi-channel, in such a scenario may get passage of 14 × 1
If an EKG, is 5 × 1, if the filter 16
may get the last data of a 10 × 16, which convolution may be a layer in the network
For the next layer convolutional network
if a 10 × 16 input data may be used for a 5-dimensional convolution filter, which requires matching channels 16
if the filter 32, the output of the further layer 6 is × 32
The 2D data, when the processing data of 10 × 10 × 16 are similar
may be used 5 × 5 × 16 convoluted, wherein two channels to match the number 16
will get a 6 × 6 is output
if the 32 is a filter, the output sources is 6 × 6 × 32, which is 32
All of these methods can also be applied to one-dimensional data, may use the same feature detectors at different locations
for example, in order to distinguish the difference heartbeat EKG signals, can use the same features in different positions to detect the heartbeat of the time axis
so convolution 1D network can likewise be used for the data, a one-dimensional data for many applications
actually use a recursive neural network processing
3.3-dimensional data
with the number of columns 1D digital matrix or different, now have a 3D block, a data input 3D
In the CT scan as an example, this is a use of X rays, and then outputs a 3D model of the body
CT scan is that it can achieve obtain different body segments (Image)
When a CT scan, and is now doing the same thing, you can see different slices torso, this data is inherently 3-dimensional
An understanding of this embodiment of the data is:
Suppose the data now includes a length, width and height, wherein each slice corresponds to a slice of the torso
如果想要在3D扫描或CT扫描中应用卷积网络进行特征识别
也可以从二维和一维的卷积里得到想法，并将其应用到3D卷积中
为了简单起见，如果有一个3D对象，比如说是14×14×14
这也是输入CT扫描的宽度与深度（后两个14）
再次提醒，正如图像不是必须以矩形呈现，3D对象也不是一定是一个完美立方体
所以长和宽可以不一样
同样CT扫描结果的长宽高也可以是不一致的
为了简化讨论，仅使用14×14×14为例
如果现在使用5×5×5过滤器进行卷积，过滤器现在也是3D的，这将会给一个10×10×10的结果输出
技术上来说也可以再×1（编号1所示），如果这有一个1的通道
这仅仅是一个3D模块，但是数据可以有不同数目的通道
那种情况下也是乘1（编号2所示），因为通道的数目必须与过滤器匹配
如果使用16过滤器处理5×5×5×1，接下来的输出将是10×10×10×16
这将成为3D数据卷积网络上的一层
如果下一层卷积使用5×5×5×16维度的过滤器再次卷积，通道数目也与往常一样匹配
如果有32个过滤器，操作也与之前相同，最终得到一个6×6×6×32的输出
某种程度上3D数据也可以使用3D卷积网络学习，这些过滤器实现的功能正是通过3D数据进行特征检测
CT医疗扫描是3D数据的一个实例
另一个数据处理的例子是可以将电影中随时间变化的不同视频切片看作是3D数据
可以将这个技术用于检测动作及人物行为
总而言之这就是1D、2D及3D数据处理，图像数据无处不在
以至于大多数卷积网络都是基于图像上的2D数据
但希望其他模型同样会有帮助