今天我们将介绍RapidMiner图像挖掘扩展的最后一个部分，如何利用该图像挖掘扩展进行生产线监控和基于图像的相似产品推荐。
Today we will introduce the last part of the RapidMiner image mining extension：how to use the image mining extension for production line monitoring, image based recommendation system ~
生产线监控
Production Line Monitoring
为什么要使用自动生产线监控?
日夜关注产品线可能非常累人
注意力不足可能会导致忽视缺陷产品的产生
机器比人稳定、快速、节约
可以监控哪些因素？
形状
内部缺陷
外部缺陷
移位
输出质量
对象计数
分析结果如何利用？
1. 结果可以存储在数据库用于长期监控
绘制统计信息
精度测试
2. 可以基于结果触发操作
移除有缺陷的产品
提醒操作员
Why to use automatic line monitoring?
Watching line of products for nights can be very tiring
Little in attention can cause an overlook of defective product
Machines are more stable than human、faster and cheaper
What can be monitored?
Shapes
Inner defects
Outer defects
Displacement
Output quality
Objects count
What can be done with results?
1. Results can be stored in a database-long term monitoring
Plotting statistics
Accuracy performance during time
2. Action can be triggered based on the results
Defective product can be removed
Operator can be alerted
RapidMiner流程概览
process overview
训练部分
Training part
读取图像Read Image
统计区域合并Statistical Region Merging
将图像基于相似性分成不同的区域或组，这个例子中，需要找到的三个区域分别是“巧克力、坚果、背景”
Images are divided into different regions or groups based on how similar the regions are. In this example, it is meant to be able to find areas which are 'chocolate/nut/background' and then label them in different colours.
标记Labeling
手动点击鼠标对不同区域进行标记
Use mouse buttons to label regions
从分割后的图像中提取特征
Extract features from segmented images
图像不能直接放入决策树模型，需先用数字描述图像中的特征，可根据不同物体的大小、方向、角度等提取数字，这些数字将与标记步骤一起放入决策树模型。
It is not possible to just put in an image into a decision tree model, so you must first make sure you have numbers describing the things that you have in the image. This example extracts numbers for the size of the different things, their direction/angle, etc. These numbers, together with the labeling step (clicking to show if it is a good cake or a manufacturing mistake) are then put into the decision tree model.
决策树Decision tree
这里使用决策树来学习与巧克力质量好坏相关的特征。尝试其他算法，如深度学习，看看是否可以实现更高的准确性。您可以使用任何支持多项式标签的模型学习算法。
A decision tree is used here to learn the features which are related to good and bad quality chocolates. Try other algorithms such as DeepLearning and see if you can achieve greater accuracy. You can use any model learner which support polynomial label.
输出结果Output
该算法可以在训练之后根据训练示例识别好的和坏的巧克力，可将其放置在生产线中实时进行分析。
The algorithm can after being trained recognize good and bad chocolates according to the training examples. It can then be put into deployed in a production line setting to do this in real time.
基于图像的产品推荐系统
Visual Market Basket Recommendation System
利用RapidMiner图像挖掘扩展进行基于图像的产品推荐，以下问题便可迎刃而解：
客户最有可能对哪些其他产品感兴趣？
客户离开的原因是什么？
有多少客户正在寻找替代品？
如何提供类似的产品？
应用方向：
购物篮数据分析，交叉销售，目录设计，销售活动分析，网页日志（点击流）分析和DNA序列分析等。
With the visual market basket recommendation system, we can answer the following question:
Which other products the customer could most likely be interested?
What is the reason why customers leave?
How many customers are looking for alternatives to the curently viewed product?
What about to provide also a similar product?
Applications：
Basket data analysis, cross-marketing, catalog design, sale campaign analysis, Web log(click stream) analysis, and DNA sequence analysis.
可通过以下两种方式实现：
This can be realized through the following two methods:
1. 频繁模式分析：找到数据的固有规律（比如同时购买了尿布和啤酒的顾客，还有可能购买哪些产品？）
Frequent Pattern Analysis：finding inherent regularities in data（e.g. for customers who buy diaper and beer at the same time, what other products the customer is likely to be interested in?）
2. 找到相似产品：视觉推荐可以显著地帮助为某产品提供有趣的替代品
Finding similar products：Visual recommendation can significantly help in providing interesting alternatives for a product
今天我们讨论的是第二种方法，通过图像挖掘找到相似产品向顾客进行推荐
Today we are talking about the second method: find similar products to help providing alternativs for products
模型训练Training model
为整个图像提取“全局特征”，与上一个例子相似，此例中三个种类分别为“其他、相似、不相似”
It gets numbers for the whole image (‘global features’). These are then put into the same kind of model as above, but with three categories: other groups/similar/not similar.
参数优化Parameter Optimizing
可以通过参数优化来增加图像相似性结果的准确性
Accuracy of the image similarity results can be improved by parameter optimization.
今天我们将结束RapidMiner图像挖掘扩展的介绍，通过这三天的了解，是否让您更能了解如何利用图像挖掘内在价值呢？
Today we will end the introduction of RapidMiner Image Mining Extensions. Doyou have better understanding on how to use the image to tap the intrinsic value?
对图像挖掘感兴趣或有任何问题都欢迎在下方留言哦~
Let us know in the comments below!
想要了解更多RapidMiner信息，您可以访问www.rapidminerchina.com或拨打电话4006-326-339咨询~
For more information on RapidMiner, you can visit www.rapidminerchina.com or call 4006-326-339.