【New Infrastructure for AI, series 2 of 3】Intelligent Image Recognition
Computer vision has been a vital science since the invention of the computer. The traditional method adopted an expert system and wrote specific algorithms to process different objects based on their characteristics. The developers must be equipped with a wealth of specialist knowledge and experience to complete the algorithm. With the rise of deep learning in recent years, various researchers began to use deep learning for computer vision processing. Many competitions proved that deep learning has better accuracy, close to the human level than the traditional expert system. For executing image recognition via deep learning, the developer only needs to collect a great deal of object image database, select a proper deep learning model for training and enter the image to be recognized into the trained model to output the recognition result.
Nuvoton's intelligent image recognition for remote meter reading integrates deep learning to execute image processing and recognition while implementing transmission via wireless methods such as Wi-Fi or NB-IoT to achieve intelligent remote meter reading. The solution supports GPIO or image capturing interface (CCAP) with an external image sensor to acquire the object before processing the image. The recognition result is sent back after inputting the item into the trained deep learning model.
The solution can recognize ten numeral images (0~9) with the Nuvoton NuMicro® M480 series microcontroller. Based on the core of Arm Cortex-M4F with an operating frequency of 192MHz, it can quickly execute the user's algorithm. Simultaneously, its floating-point unit (FPU) and digital signal processor (DSP) can significantly improve the neural network algorithm's execution efficiency in machine learning. The solution is available to support a recognition rate of 320×240 with an image frame reaching 30 FPS due to the efficient core processing speed, flash reaching a maximum of 512 KB, large capacity memory of built-in SRAM up to 160KB, and hardware camera interface. The built-in EBI interface x80 or SPI interface can connect the external LCD screen to display the image. The SDHC interface can connect the SD card to save photos and related information. Also, the multi-path UART interface can connect different wireless pass-through modules at the same time to achieve the function of remote transmission.
M480 series (Cortex-M4) development platform and product features:
- Core speed up to 192MHz with sufficient memory space of 512KB Flash/160KB SRAM to support the system design requirement
- Arm Cortex-M4Fwith built-in FPU and DSP
- Hardware image capturing interface can support CCIR656/601 transmission
- Support 2 SD card storage devices (up to 2x 32GB)
- Support 5 SPI and external SPI Flash for metafile saving
- Support EBI interface, external SRAM storage (up to 3x 1MB), or support i80 LCD screen
- SPI M interface with cache memory function to support the code executing in external flash
Suggested microcontroller:
- M481 Base Series
- M482 USB FS OTG Series
- M483 CAN Series
- M484 USB HS OTG Series
- M485 Crypto Series
- M487 Ethernet Series
Matching development board:
Application link:
https://www.nuvoton.com/applications/industrial/smart-water-meter/