Artificial intelligent sorting machine which selects cucumber by deep learning is being developed
Speaking of the most gorgeous computer technology at the moment, it can be said that artificial intelligence using deep learning is true, but such waves seem to be coming to the field of agriculture as well. Makoto Koike, a former control system developer and currently running cucumber farmers who are family business,Using Google's TensorFlowBy utilizing the learned AI, we are developing a device that automatically sorts harvested cucumbers and classifies them according to grades.
[CCB 9] Cucumber sorting machine Prototype 2 machine | Workpiles
http://workpiles.com/2016/08/ccb9-prototype2-complete/
Google Cloud Platform Japan Official Blog: TensorFlow connecting cucumber farmers and deep learning
http://googlecloudplatform-japan.blogspot.jp/2016/08/tensorflow_5.html
The "cucumber sorting machine" developed by Mr. Koike has a structure in which a sorting belt conveyor is mounted beside the vertical type rack. When cucumber is placed on the right rack, the camera shoots the exterior, AI discriminates the grade, and it is put in a predetermined cardboard box with a belt conveyor.
It looks like the cucumber sorter is actually in operation. In the past you can see how the machine automatically performs the work that you were doing with human eyes and hands.
TensorFlow powered cucumber sorter by Makoto Koike - YouTube
When you put the cucumber on the tray of the sorter, ...
Based on the appearance of cucumbers taken with multiple cameras, AI judges which classes to sort. It seems that AI has shown pictures of thousands of cucumbers beforehand and learns the ability to judge grades by themselves.
When the judgment is over, the tray on which the cucumber was placed collapsed with the cattle and the cucumbers were dropped on the belt conveyor ......
A belt conveyor sends cucumber to the left.
And the lever moves and shoots the cucumber into the cardboard. The corrugated board seems to be placed so that the grade goes higher as it goes to the left, and the cucumber that was judged earlier was B grade goods, but since the viewer left on the belt conveyor, the splendid cucumber is placed in a box of a higher grade It seems to be put in.
Koike told me that Google's artificial intelligence "Alpha Go (Alpha Go)" looked like going to be equal with the world's strongest players and got instinctively that "This is a terrible thing happening" He said he had started developing a cucumber sorting machine using deep learning.
Mr. Koike says that her parents are running cucumber farmers, in the busy season I will see mothers sorting cucumber for 8 hours a day. Koike, who thought that if the sorting work was automated, Mr. Koike thought that time and effort would be directed towards the original cucumber making, he seems to have completed the completion of the prototype in just four months. It is said that Unit 2 of the prototype was completed at the time of writing the article.
Mr. Koike's father, who believes in making fresh cucumbers, has a commitment to "cucumbers stinging thorns" and that cucumbers with irregularities of irregularities remaining on the surface are considered witnesses of quality.
Actually, cucumber grade does not have such standards, it seems that farmers make their own judgment. In Koike's place, we are sorting out 9th grade by cucumber weight, color, gloss, scratches, etc. This is not a job that humans can easily do.
AI using deep learning is used as a technology to automate such work.
Koike says that he had not had the opportunity to contact deep learning and machine learning before, since he uses "TensorFlow" which is an open source library, he seems to have entered the world of deep learning. Although TensorFlow is not familiar with mathematical models and optimization algorithms necessary for implementing deep learning, it is possible to "start touching for the time being" by using open samples and tutorials, and as a tool for getting started It can be said to be the best one.
The system configuration of Prototype 2 is as follows. Raspberry Pi 3 controls image capture by Web camera, and judges the presence or absence of cucumber by small scale neural net by TensorFlow. Then, we send the taken cucumber image to the recognition engine based on deep learning which runs on TensorFlow on Linux machine. For cucumber image recognition, TensorFlow sample code "Deep MNIST for Experts"Based on the base, we use a slightly modified deep neural network. It seems to change the resolution and the classification number of image to be handled for cucumber sorting with the configuration that the congruent layer and the pooling layer are passed several times and then the entire binding layer is arranged.
It seems that it takes several seconds per one judgment speed at the present time. In addition, it seems that learning is in the early stages, the correct answer rate of judgment is about 70%, and there seems to be room for improvement. However, these tasks are expected to make dramatic improvements by learning more cucumber by distributed learning by the cloud.
It used to be a technology developed by farmers to alleviate the serious agricultural workAgricultural equipment manufacturers such as Yanmarwas. About a century later, agriculture in the 21st century may be due to artificial intelligence using AlphaGo 's servers below to reduce burden and increase efficiency.
In addition, details of the actual technology etc. can be read with the original source at the top of the page. Also, Mr. Koike's cucumber sorting machine has been posted on Google's English blog and introduced to the world.
How a Japanese cucumber farmer is using deep learning and TensorFlow | Google Cloud Big Data and Machine Learning Blog | Google Cloud Platform
https://cloud.google.com/blog/big-data/2016/08/how-a-japanese-cucumber-farmer-is-using-deep-learning-and-tensorflow
Related Posts: