Fathers and elders , Villagers ！ You know faces 、 goods 、 Vehicle identification , Search for pictures with pictures, even automatic driving , What's the technology behind it ？
At first, Xiaobian thought it was image classification 、 Target detection, these things , What's the difficulty ？ But those who are proficient in these applications BAT Senior engineers make millions a year , I'm proud of my life ！！！
When Xiaobian is turning into lemon essence ,BAT The great God said ： Behind this is the comprehensive use of target detection 、 Image classification 、 Measure learning 、 The key to image retrieval is 【 Universal image recognition system 】…
What is measurement learning ？ What is image retrieval ？ What is a universal image recognition system ？！ It seems that it's still me Too Simple,Too Naive 了 …
I really can't make a million a year ？！ When Xiaobian beat his chest and feet , Xiaobian suddenly found a general image recognition system to quickly build artifact ！OMG！ This dream will come true soon ！
hurriedly Star Collect ：https://github.com/PaddlePaddle/PaddleClas
So what's so amazing about this project , Image recognition is better than image classification 、 Where is the strength of target detection ？
take [ Commodity identification ] Take a chestnut , If you use pure image classification and object detection , You'll find that ：
Tens of thousands of commodity categories ： There's no way to put all the categories into the training set in advance ！ The training set is not complete, how to train the algorithm ？
The sample categories are extremely uneven ： The quantity distribution of each kind of goods is uneven , Some categories are just 1、2 A picture ！ Even if such a class is put into algorithm training , The recognition accuracy is also very, very low .
Category update is very fast ： Various businesses keep introducing new categories , Every time you add a new product, you have to work hard to retrain the model ！
And using image recognition , Not only can the above problems be solved perfectly ！ And it's very fast , Simple and easy to understand . When you have a picture of the object you want to recognize , Just three steps , Multiple categories 、 Small sample 、 Data imbalance is no longer a problem ！ And it's not just product identification , It's also possible to carry out vehicle 、 Face 、Logo、 Pedestrian identification ！！！ Let's identify what we want to identify ！（ It's the feeling of freedom. That's right ！）
And this image recognition system 4 There are three core components , It's all carefully polished . Whether it's used alone or in series , They all have extraordinary effects ：
Subject detection ： With high precision and ultra light weight PP-YOLOv2 Detection algorithm , Fast subject detection of image , Improve recognition efficiency .
Backbone network ： selected 6 Series Backbone, Covering the most sophisticated mobile model and high-precision server model , Support the rapid modification of the structure , Meet the needs of different scenarios .
Measure learning ： Integrate ArcMargin, CenterLoss, TriHard And other leading measurement learning methods in the industry , And can be combined at will , It's easy to train robust image features .
Search system ： Integrated Baidu self-developed Möbius Algorithm , Efficient vector retrieval , And can update the search database at any time , Long term use of one training .
Developers can not only assemble and use these four modules individually or independently , You can also use the built vehicle identification directly 、LOGO distinguish 、 Commodity identification 、 There are four systems for animation recognition . Just supplement the search database , It can be put into operation directly ！
With their help , Keep up with the programmers of big factories , A million a year is not a dream ！（ Gradually showing a rich smile ~ ）
So strong 、 Project with heart , What are you waiting for ？！ Not in a hurry Star Let's get in the car ！
Quick experience ：