10 Convolutional Neural Networks for Deep Learning 2

本文是接着上一篇深度学习之 10 卷积神经网络1_水w的博客-CSDN博客

目录

1 出现原因 

2 General structural framework

（1）General structural framework：卷积层 ——Extract features using convolution kernels

The essence of the convolution kernel：

总结

特征图可视化

（2）General structural framework：池化层

（3）General structural framework：全连接层

CNNThe process by which a convolutional neural network recognizes images 

3  What Convolutional Neural Networks are good at

---

1 出现原因 

在 CNN 出现之前,图像对于人工智能来说是一个难题,有2个原因：

        * 图像在数字化的过程中很难保留原有的特征,导致图像处理的准确率不高;

        * 图像需要处理的数据量太大,导致成本很高,效率很低;

 

  When working with image data on a fully connected network,拉伸成1维：The local spatial features of the image are lost.

CNN 解决了这个问题：他用类似视觉的方式保留了图像的特征,当图像做翻转、旋转或者变换位置时,它也能有效的识别出来是类似的图像.

2 General structural framework

（1）General structural framework：卷积层 ——Extract features using convolution kernels

 

The essence of the convolution kernel：

The first layer extracts edges,The second layer combines the extracted edges into shapes,The third layer extracts obvious objects.

• A single convolution kernel represents some kind of simple feature of the image.比如垂直边缘、水平边缘、颜色、纹理等.
• All convolution kernels add up as a collection of feature extractors for the entire image.
• Stacking multiple layers of convolution can progressively extract higher levels、更复杂、更抽象、More generalized features.

卷积核深度（通道数）should match the input！

 

*Use convolution kernel for convolution calculation,

* 卷积结果,

 * 卷积结果：Consider multiple filters

 *因此,如果我们有6个5x5x3的滤波器,我们将获得6a separate activation map：

 We stack them,以获得尺寸为28x28x6的“新图片”！

The size of the output of the convolutional layer？ --- 通道数

         * The number of channels of the convolution kernel used by a convolutional layer is equal to the number of channels of the input data of this layer

         * The number of output channels of a convolutional layer is equal to the number of convolution kernels used in that layer

 

The size of the output of the convolutional layer？ --- 长/宽

在实际中：Commonly used zero-pad borders,保证输入输出尺寸相同 .   (N – K+2P) / stride + 1

通常的情况是,The convolutional layer stride is set to 1,滤波器大小为KxK,则一般使用（K-1）/ 2pixels are zero-padded（The size will be preserved in space）

eg：输入图片大小: 32x32x3 ,10个卷积核,大小为 5x5x3,步长为 1, pad = 2.

（1）输出图片大小: (32-5+2*2)/1+1 = 32 ,因此输出大小为 32x32x10

（2）What is the number of parameters in this layer？ Each convolution kernel has 5 * 5 * 3 +1 = 76个参数 (+1 for bias) => 76 * 10 = 760

 

总结

Given a convolutional layer,

需要四个参数：

        * 卷积核数目F

        * 卷积核大小K

        * 步长S

        * Number of zero paddingP

输入图片大小为 𝑊1 × 𝐻1 × 𝐷1;

The output size after convolution is 𝑊2 × 𝐻2 × 𝐷2 ,其中

        * 宽度：𝑊2 = (𝑊1 − 𝐾 + 2𝑃)Τ𝑆 + 1

        * 高度：𝐻2 = (𝐻1 − 𝐾 + 2𝑃)Τ𝑆 + 1

        *通道数=卷积核数： 𝐷2 = F

通过参数共享,Each filter is introduced𝐾 ∗ 𝐾 ∗ 𝐷1 + 1 个参数,因此一共有 𝐹 ∗ 𝐾 ∗ 𝐾 ∗ 𝐷1 + F个参数;

输出结果中,第𝑑个切片（大小为 𝑊2 × 𝐻2）is the pass on the input𝑑convolution kernel with stride𝑆Do convolution and then add 上𝑏𝑖𝑎𝑠的结果.

特征图可视化

• 通过deconvolution,可以对feature map 进行可视化;
• 浅层layer学到的特征为简单的边缘、角点、 纹理、几何形状、表面等;
• 深层layer学到的特征则更为复杂抽象;
• 人工只能胜任简单卷积核的设计,如边缘;
• The convolution kernel weights of each layer of the convolutional neural network are driven by dataMove to learn,不是人工设计的.
• Data-driven convolutional neural networks逐层Learn from simple to complex features（模式）;
• 复杂模式是由简单模式组合而成;
• 不同的边缘->不同纹理->Different geometric figures->different silhouettes->不同的物体...
• The combination of shallow patterns can be varied,The patterns that make the deep layer descriptive can also vary,So have a strong table

Ability to reach.

---

（2）General structural framework：池化层

 池化层（下采样）——数据降维,避免过拟合

 

（3）General structural framework：全连接层

CNNThe process by which a convolutional neural network recognizes images 

 

3  What Convolutional Neural Networks are good at

以（多维）数组形式signal appearing

Strong local correlation的信号

A signal that features can appear anywhere

Object translation and deformation invariant signals

  1D Convolutional Networks：Sequential signal,文本

        • 文本、音乐、音频、演讲、时间序列

  2D Convolutional Networks：图像,时频表示（Voice and Audio）

        • 目标检测、定位、识别

  3D Convolutional Networks：视频,体积图像,断层扫描图像

        • 视频识别/理解

        • Biomedical Image Analysis

        • Hyperspectral image analysis