【CV-Learning】Convolutional Neural Network

学习路线

————————————————
首先,应该具有Fully connected neural network foundation后,to study this article.
————————————————

图像表示

像素表示：Directly utilize raw pixels as features,展开为列向量.
Ps：CIFAR10Each image of the dataset is（32323）3072维向量.

分类模型

全连接神经网络的瓶颈

例：图像尺寸为32323,The number of weights for each neuron in the hidden layer is 3072+1=3073.
例：图像尺寸为2002003,The number of weights for each neuron in the hidden layer is 120000+1=120001.

问：为什么要加上1？
答：When calculating the number of weights for each neuron in the hidden layer,Except that every connecting edge is considered,There is also an offset to considerb.

由此可见,When the image size increases,The number of weights for each neuron is greatly increased,If we consider multiple neurons on a hidden layer and multiple hidden layers,The amount of computation will be quite large.

结论：Fully connected neural networks are only suitable for processing small images.

卷积神经网络

Since the fully connected network is suitable for processing small images,That is, the dimension of the input layer is small.From this, we can first use the convolution kernel to process the image,For example, it will be obtained from the preparatory knowledge in the previous section48dimensional feature vectors are processed as input,After that, the fully connected network processing is performed.

卷积层

基于卷积核组的图像表示

卷积核（Slightly different from those in the convolution kernel group）

1.Not just width and height,还有深度.Often written in the following form：宽度 * 高度 * 深度.
2.The convolution kernel parameters include not only the weights stored in the kernel,Also includes an offset value.

卷积操作

The depth of the convolution kernel of the next layer of convolution operation should be consistent with the depth of the corresponding layer of the feature,That is, the number of convolution kernels of the previous layer of convolution operation should be the same.

卷积步长

卷积步长（stride）：卷积神经网络中,卷积核可以按照指定的间隔进行卷积操作,This interval is the convolution stride.
The larger the convolution stride,The smaller the size of the output feature map group.

边界填充

To prevent after multiple convolution operations,The output image is getting smaller and smaller,We need to do boundary padding to reach the input、The output image size is required to be consistent.

Feature corresponding graph size calculation

激活层

The convolutional neural network used in this paperReLUfunction to activate.

池化层

池化的作用：对每一个特征响应图独立进行,降低特征响应图组中每个特征响应图的宽度和高度,减少后续卷积层的参数的数量,降低计算资源耗费,进而控制过拟合;Also makes the convolution kernel see more things,Extract a rougher image（Equivalent to a large variance effect）.

池化操作：对特征响应图某个区域进行池化就是在该区域上指定一个值来代表整个区域.
The pooling operation is performed independently for each feature response map.

常见的池化操作：
1.最大池化——使用区域内的最大值来代表这个区域
类似于非极大值抑制.

2.平均池化——采用区域内所有值的均值作为代表

池化层的超参数：池化窗口和池化步长.

损失函数

交叉熵损失：Measures the distance between the classifier's predicted distribution and the true distribution.
两个分布越接近,The smaller the loss of cross entropy.

优化算法

1.SGD、带动量的SGD

Momentum method advantage：due to the presence of momentum,The algorithm can rush out of a local nadir or saddle point,找到最优解.

2.ADAM

图像增强

图像增强

存在的问题：过拟合的原因是学习样本太少,导致无法训练出能够泛化到新数据的模型.

数据增强：是从现有的训练样本中生成更多的训练数据,其方法是利用多种能够生成可信图像的随机变换来增加样本.
The goal of data augmentation：模型在训练时不会两次查看完全相同的图像.这让模型能够观察到数据的更多内容,从而具有更好的泛化能力

增强方法：
1.翻转
2.随机缩放&抠图

3.色彩抖动

4.平移
5.旋转
6.拉伸
7.径向畸变
8.裁剪
…

北京邮电大学–计算机视觉与深度学习–鲁鹏