What should be a good cucumber? Google AI tells you |

Makoto Koike was originally an engineer at a Japanese car company. About a year ago, he resigned from his job and returned to his hometown to help his parents run a cucumber farm. The farm was not large, but the work of classifying cucumbers caused Makoto to suffer a lot.

Every farm in Japan has different classification standards for cucumbers, which is not as simple as we think of. In Makoto's own farm alone, there are as many as nine kinds of cucumbers of the same species. In general, bright colors, thorns, and symmetry are good melons.

The following is an impressive chart of the Makoto family's 9 cucumbers, which are descending from top to bottom.

Just picked a cucumber in your hand, you have to carefully observe its length, thickness, color, texture, whether there are small scratches, curved or straight, not much thorn ... ... to meet the 9 standards Correspondence to see which level it belongs to is not an easy task to learn.

It takes several months for a person to master the entire classification system. Therefore, at the picking season, Makoto’s home is too busy to hire someone for help.

However, Makoto always believed that the classification of cucumbers should not be the main job of melon farmers. The most important task for melon farmers should be to concentrate on producing delicious cucumbers. So he decided to give the sorting work to the machine, but the cucumber classifier on the market was either poor or too expensive and not suitable for small farms.

At this time, he saw the Alpha Dog Go tournament, which was immediately attracted by artificial intelligence, developed the idea of ​​using machine learning to create a cucumber sorter, and began to study the Google open source TensorFlow platform.

It is worth mentioning that the use of TensorFlow does not require advanced mathematics models, optimization algorithms and other professional knowledge. You just need to download a simple code and then read the tutorial to start working.

This is the scene of Makoto's cucumber sorter work: if a cucumber belongs to a certain category, the small brush will push it into the corresponding box.

Makoto uses the Raspberry Pi 3 as the main controller and is equipped with a camera to take pictures. The photos were passed on to the TensorFlow platform and were initially run on a small neural network to determine if it was a cucumber. Afterwards, the photos that have been determined to be cucumbers are then transmitted to a larger Linux server-based neural network to classify cucumbers according to different traits.

The following is a system diagram of a cucumber sorter:

First of all, machine learning requires a database. In order to train the model, Makoto spent 3 months feeding 7000 cucumber photos. These photos were tagged by Makoto's mother.

The accuracy of Makoto's test was very high, reaching 95%, but when he really used this system for practice, the recognition accuracy rate dropped to 70% at once. Makoto suspects that this neural network model has an "overfitting" problem, which is a phenomenon that occurs due to the lack of database data.

In addition, the amount of computation required for deep learning is large, and Makoto uses a typical Windows PC to train the neural network, which is inefficient. Although he had previously reduced all photos to 80 x 80 pixels, the system still takes 2-3 days to complete 7000 photos.

The result of this low-resolution photo is that the system does not currently recognize colors, textures, scratches, and small stings. It can only tell the shape, length, and whether it is curved. If you want to increase the resolution of your photos, the amount of calculations in the system will soar and the efficiency will slow down.

So Makoto is currently planning to use Google’s Cloud Machine Learning platform to further improve his cucumber sorter.

Via Engadget

Extended reading:

Secret TensorFlow: Google open source in the end is what?

Climb the top in the world! How does AlphaGo compare to other AIs?

Press Brake Protection

BLPS laser safety protective device is designed for personal safety used on hydraulic bender.
The dynamic test technology it used has passed the Type 4 functional safety assessment by TUV, and get the national invention patent. The product reaches the advanced technological level of similar products.
BLPS laser safety device provides protection zone near the die tip of the bender to protect fingers and arms of the operator in close to the upper mold die tip. It is the most effective solution so far to preserves the safety and productivity of the bender.

Press Brake Protection,Laser Guarding Device,Press Brake Guarding Systems,Press Brake Guarding

Jining KeLi Photoelectronic Industrial Co.,Ltd , https://www.sdkelien.com