Fleeting Day（VOL.11）

### R-CNN

#### Non-Maximum Suppression

1. 从最大概率矩形框F开始，分别判断A~E与F的重叠度IOU是否大于某个设定的阈值;
2. 假设B、D与F的重叠度超过阈值，那么就扔掉B、D；并标记第一个矩形框F，是我们保留下来的。
3. 从剩下的矩形框A、C、E中，选择概率最大的E，然后判断E与A、C的重叠度，重叠度大于一定的阈值，那么就扔掉；并标记E是我们保留下来的第二个矩形框。
4. 就这样一直重复，找到所有被保留下来的矩形框。

Selective Search for Object Detection (C++ / Python)

SS算法根据颜色，纹理，大小和形状兼容性，计算相似区域的层次分组。（It is based on computing hierarchical grouping of similar regions based on color, texture, size and shape compatibility.）

1. 生成区域集R
2. 计算区域集R里每个相邻区域的相似度S={s1,s2,…} (颜色、纹理、尺寸和空间)
3. 找出相似度最高的两个区域，将其合并为新集，添加进R
4. 从S中移除所有与step3中有关的子集
5. 计算新集与所有子集的相似度
6. 跳至step3，直至S为空

#### SPP

Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

15年何恺明等人在R-CNN的基础上提出的SPP-CNN中，在卷积层和全连接层之间加入空间金字塔池化结构（Spatial Pyramid Pooling），从而代替原R-CNN需要在输入CNN前对各个候选区域进行剪裁、缩放操作使其图像子块尺寸一致的做法，有效避免了R-CNN算法对图像区域剪裁、缩放操作导致的图像物体剪裁不全以及形状扭曲等问题，也解决了卷积神经网络对图像重复特征提取的问题，大大提高了产生候选框的速度，且节省了计算成本。

CNN中，卷积层和池化层可以对任意大小的图像卷积，只有全连接层对图像尺寸大小有要求，因其输入向量的维度是固定的。为使得对任意大小的输入，经过卷积层后输入全连接层的feature maps的维度都是相同的，对最后一个卷积层的输出做三种步幅、窗口大小的池化处理，分别得到大小为$4\times4$、$2\times2$、$1\times1$的三个feature maps，再将这三副图依次排列，可以组成大小为$(4\times4 + 2\times2 + 1\times1)times256 = 5376$的特征向量（$256$为通道数）。这样不管输入图像尺寸多大，经过卷积层后输入全连接层的大小都是$5376$，从而实现对不同大小的输入通过相同网络进行分类。

SPP-CNN的大体流程如下：

1. 和R-CNN一样，用SS算法，对待检测的图片进行搜索出2000个候选窗口；

2. 和R-CNN一样，用卷积神经网络进行特征提取，但是SPP-Net用到了SPP。具体操作为：把整张待检图片输入CNN中，进行一次性特征提取得到feature maps，然后在feature maps中找到各个候选框的区域，再对各个候选框采用金字塔空间池化，提取出固定长度的特征向量。因为SPP-Net只需要一次对整张图片进行特征提取，相比R-CNN中要输入每个候选框，速度是大大加快。

3. 最后也和R-CNN一样，采用SVM算法进行特征向量分类识别。

### Economist 190427

Screening for lung cancer is a controversial idea

1. Introducing the topic of the lung cancer by Mr Thomas’s case
2. Lung cancer is the cancer that kills the most
• Three-quarters of people have already arrived at Stage IV at the moment of diagnosis
• A lot of places have resisted extending the screening programmes to cancers of the lung
3. The question is whether such screening would do more good than harm
• Finding the prostate cancer out involves tests such as biopsies which are invasive, painful and may cause infection
• the tumour would not have shortened a patient’s life and the men who are offered treatment risking side-effects
4. NIST’s results shows screening for lung cancer might be beneficial
• In one part of the world that conviction of lung cancer might be beneficial is being translated into action
• trivial 微不足道的
• jargon 术语
• infantial 幼儿的
• side-effect 副作用
• impotence *无能
• computerised tomography (CT)
• eminently 极其，特别
• nuanced 细微的
• sinister 凶险的

Finding tumours earlier would permit them to be treated before they spread—improving outcomes and reducing medical bills.

Yet a lot of places which happily invite people to participate in screening programmes intended to uncover cancers of the breast, bowel, prostate gland and cervix have resisted extending the idea to cancers of the lung.

Though some sceptics still argue that money spent on screening would be better deployed trying to stop people smoking—or, better still, preventing them from starting in the first place—the case for screening smokers for lung cancer now seems a good one.