列联表

频数列联表

频率列联表

Likehood Funciton

似然函数

定义

记已知的样本集为：

$$D=\left\{x_{1}, x_{2}, \cdots, x_{N}\right\}$$

联合概率密度函数$p(D | \theta)$称为相对于$\left\{x_{1}, x_{2}, \cdots, x_{N}\right\}$的θ的似然函数:

$$l(\theta)=p(D | \theta)=p\left(x_{1}, x_{2}, \cdots, x_{N} | \theta\right)=\prod_{i=1}^{N} p\left(x_{i} | \theta\right)$$

如果$\hat{\theta}$是使似然函数$l(\theta)$最大的$\theta$值，则$\hat{\theta}$应该是“最可能”的参数值，那么$\hat{\theta}$就是$l(\theta)$的极大似然估计量。

它是样本集的函数，记作：

$$\hat{\theta}=d\left(x_{1}, x_{2}, \cdots, x_{N}\right)=d(D)$$

$\hat{\theta}\left(x_{1}, x_{2}, \cdots, x_{N}\right)$称作极大似然函数估计值

Gesture recognition by Deep Neural Network

Previously on Camshift

In the last TP, we used the camshift algorithme to track the hand and take picture of it. Then we resize the picture to 16X16 matrix and store it as the data set to train our neural network.
I spend an hour to capture the gesture of my hand and I only got like 1500 pictures. So in order to have more data to train the network. I randomly select 1000 pictures I have and rotate each of them with a random angle and add them into the data.(The code for selecting and rotating them is here). So now I have a data set with 2500 pictures.

We have four letters to predict C,V,I and O
We will use two kinds of Network, MLP and CNN

直方图和直方图均衡化(Histogram Equalization)

图像直方(Image Histiograms)

灰度直方图是对每一阶灰度的像素数量进行计数。
像素数量常常可以改用像素的概率来表现： Pk=nk/N 其中k为灰阶，n为k灰阶像素的数量，N为像素数量总数

OpenCV 反向投影

1.什么是反向投影

在OpenCV的中文文档上反向投影的描述是这样的：

• 反向投影是一种记录给定图像中的像素点如何适应直方图模型的像素分布的方式。
• 简单的讲， 所谓反向投影就是首先计算某一特征的直方图模型，然后使用模型去寻找图像中存在的该特征。
• 例如， 你有一个肤色直方图 ( Hue-Saturation 直方图 )，你可以用它来寻找图像中的肤色区域

2.反向投影的步骤

从上面文档的描述中可以看到，反向投影的步骤三步：
1.获取目标图像区域的直方图模型
2.将新图像中的每一个像素点与直方图进行比较，获得该像素点的颜色在直方图中的概率。
3.将结果储存到新的图像中

Lab 2: Semantic segmentation with Fully Convolutional Networks

Introduction

Semantic segmentation 语义分割

In the previous lab, we used Convolutional Networks to classify images into different classes (numbers, cats/dogs).

语义分割是一项更进一步的计算机视觉任务，其目的是给每一个像素进行分类

Semantic segmentation is a computer vision task that takes classification a step further, by assigning a class to every pixel of an image.

Of course, this problem is much harder than classification. Semantic segmentation is useful for any application that requires a dense classification, such as urban mapping, medical image segmentation, self-driving cars…

Lab 1: Convolutional Neural Networks

In this first Lab, we discover a Deep Learning framework (Pytorch), which we use to create our very first CNN (LeNet) and use it to perform handwritten character recognition.

This Lab assumes that you are familiar with the Python language. If you’re not, please do Lab 0 first: https://colab.research.google.com/drive/16XlCqmmUQvwBD3D5u0lOy1rUFfDDZdYi

Introduction

What’s a CNN?

Convolutional Neural Networks are a subclass of Neural Networks that use Convolutional layers. These layers are basically sliding filters and work quite well for vision tasks.