RKNN (Rockchip Neural Network) is a high-performance deep learning inference framework developed by Rockchip, specifically optimized for embedded devices and edge computing scenarios. It enables the conversion of models trained in mainstream frameworks (e.g., TensorFlow, PyTorch, Caffe) into a specialized format that runs efficiently on Rockchip SoCs with NPU acceleration (such as RK3588, RK3566 series). Leveraging hardware-level acceleration and model quantization techniques, RKNN dramatically improves the inference speed and energy efficiency of AI models on embedded systems. It is widely used in AIoT applications like smart cameras, robotics, and industrial inspection. more details see at GitHub - airockchip/rknn-toolkit2
rknn4Delphi wraps the RKNN C++ API interfaces, enabling Delphi developers to directly utilize the RKNN inference framework within Delphi. Since Delphi does not support ARM Linux, rknn4Delphi is currently available exclusively for Android platforms. By bypassing JNI and directly encapsulating native interfaces, the library achieves higher performance. Deployment follows the standard Delphi Android application workflow, requiring no additional steps beyond standard Delphi Android project configurations.
Prerequisites for using rknn4Delphi:
Basic knowledge of deep learning is required, and models must be converted to the RKNN format. For model conversion guidance, refer to the official repository: GitHub - airockchip/rknn_model_zoo.
Compatibility:
rknn4Delphi has been tested and validated on RK3588, RK3576, and RK3566 chips using INT8 quantization. For benchmark results of inference speeds, please refer to the table below:
Add
.\source\rknn_api.pas
.\source\rknnBase.pas
.\source\rknnClassification.pas
.\source\rknnDetection.pas
to your project, reference the example project .\example\_rknnTestAndroid
Note: The models used in the sample programs are optimized for RK3588 chips. To run these models on other chip models (e.g., RK3566, RK3576), you will need to convert the models accordingly for your target hardware.
欢迎Delphi的老人们一起来交流，加入 Delphi玩转AI 群一起来变年轻
