Background introduction

Target detection and recognition is a key technology in the field of computer vision , With the development of deep learning technology , The application scenarios of target detection and recognition are also more and more extensive . At present , There are mainly the following application scenarios :

Security area ： fingerprint identification 、 Object recognition, etc .

Transportation field ： License plate number recognition 、 unmanned 、 Traffic sign recognition, etc .

The medical field ： electrocardiogram 、B super 、 Health management 、 Nutrition, etc .

Life field ： Smart home 、 Intelligent shopping 、 Intelligent skin measurement .

However, at present, AI application development faces a long development cycle 、AI The understanding cost of software stack is high 、 The combination of algorithm model and business is difficult 、 High threshold for developer skills . In order to reduce AI The threshold of application development , Rise CANN Open source a high-performance one-stop solution for general target detection and recognition , Through its powerful customizability 、 Extensibility , For the purpose of AI Developers offer better programming options .

hot tip , If you have the following knowledge reserves , Will help to learn ：

1. have C&C++ Programming experience .

2. Understand heterogeneous computing architecture CANN It's rising AI Position and function in the whole stack .

3. Understand the application programming framework AscendCL Key features of , And can be based on AscendCL Interface development is simple AI application .

The goal is

1. Understand the functions and characteristics of the one-stop solution for general target detection and recognition

2. understand ACLlite Background and interface usage

3. Deeply understand the implementation process of the one-stop solution for general target detection and recognition

4. Be able to customize your own based on this scheme AI application

Introduction to the one-stop scheme of target detection and recognition

Click here detect_and_classify, You can view the source code of the scheme .

The overall characteristics of the scheme are summarized as follows ：

1. Support multi format input and output

The one-stop solution of general target detection and recognition supports pictures 、 Offline video 、RTSP Multiple input formats such as video stream , Based on this scheme, developers can recognize objects in different formats such as pictures and videos . In addition, in the aspect of result display , Support pictures 、 Offline video 、Web Front end and other forms of display , Developers can flexibly present recognition results according to business scenarios .

2. Support easy replacement and concatenation models

Currently, the scheme is YoloV3 Image detection model and CNN Concatenation of color classification models , It can realize basic vehicle detection and vehicle color recognition , Developers can easily modify the program code , Replace... By yourself / increase / Delete AI Model , Achieve more AI function .

3. Support efficient data preprocessing

picture 、 Video and other data are raw materials for target detection and recognition , Putting data into AI Before algorithm or model , We need to preprocess the data , In order to achieve more efficient and accurate calculation . The sample uses an independent data preprocessing module , Support developers to customize on demand , Efficient decoding 、 Cutout 、 The zoom 、 Color gamut conversion and other common data processing functions .

4. Number of pictures supported 、 Variable resolution scene customization

In the field of target detection and recognition , Developers need to deal with differences in input data formats and so on , You will often encounter the number of pictures 、 Scenes with uncertain resolution , This is also one of the most troublesome problems . such as , In the process of target detection and recognition , Because the number of detected targets is not fixed , Cause the program to wait until the pictures are saved to a fixed number AI Calculation , Wasted a lot of valuable AI Computing resources . This example opens up a convenient custom entry , Support setting multiple data volumes Batch gear 、 Multiple resolution gears , When reasoning, match flexibly according to the actual input , It not only broadens the business scenario , Save computing resources more effectively , Greatly enhance AI Computational efficiency .

5. Support multi-channel and multi thread high-performance programming

In order to further improve the flexibility of programming , Meet developers to achieve high performance AI application , This sample supports adjusting the number of threads and devices in a very friendly and convenient way , Greatly reduce learning costs , Improve the utilization of equipment resources .

6. Efficient post-processing calculation

besides , This example will be followed up by the original need in CPU Push the function of processing on to shengteng AI On the processor , Take advantage of AI The powerful computing power of the processor realizes the acceleration of post-processing , Further improve the whole AI The computational efficiency of the application .