Renesas RZ/V2L Pretrained AI Model

Overview

The RZ/V2L MPU includes a Cortex-A55 (1.2 GHz) CPU and an integrated AI accelerator, DRP-AI, designed to improve machine vision processing. DRP-AI is composed of DRP and AI-MAC. The device also provides a 16-bit DDR3L/DDR4 memory interface, an integrated Arm Mali-G31 3D graphics engine, and an H.264 video codec.

DRP-AI Features

DRP-AI delivers power efficiency that can allow operation without additional cooling such as heatsinks or fans. AI inference can be implemented in consumer and industrial devices, including point-of-sale terminals. DRP-AI supports real-time AI inference and image processing, including camera functions such as color correction and denoising, which can remove the need for an external image signal processor.

Platform Compatibility

RZ/V2L is package- and pin-compatible with RZ/G2L, enabling RZ/G2L users to upgrade to RZ/V2L for additional AI capabilities without modifying system configurations, helping to keep migration costs low.

Pretrained Plant Leaf Disease Classification Model

A pretrained plant leaf disease classification model is provided for RZ/V2L. The model classifies 38 different leaf disease or healthy conditions across 14 plant species.

Supported Plants and Leaf Conditions

Runtime Modes

The application supports three runtime input modes:

1. MIPI CSI camera input
2. Image file input
3. Video file input

Example Output

The classification result, inference time (ms), and score (%) are displayed in the upper-left corner. Frames per second (FPS) is shown in the upper-right corner.

Hardware and Software Requirements

Building the Sample Application

Building is optional if you use the provided prebuilt binaries. Before compiling, ensure the RZ/V2L AI SDK is prepared.

1. Clone the source code to the local machine
2. Start Docker
3. Mount the data directory to the rzv2l_ai_sdk_container
4. Enter the sample source directory
5. Compile the source

mkdir -p build && cd build cmake -DCMAKE_TOOLCHAIN_FILE=./toolchain/runtime.cmake .. make -j$(nproc)

The application plant_leaf_disease_classify should appear in src/build.

Application Deployment

Copy the following files to /home/root/tvm on the SD card:

• All files from the sample executable directory
• If the application was modified, copy plant_leaf_disease_classify from src/build

Running the Sample Application

When the input is a MIPI Coral camera:

cd /home/root/tvm ./plant_leaf_disease_classify CAMERA

When using a static image as input:

cd /home/root/tvm ./plant_leaf_disease_classify IMAGE sampleimg.jpg

When using a video file as input:

cd /home/root/tvm ./plant_leaf_disease_classify VIDEO plantvid.mp4

Stopping the Application

Press the Esc key to exit the application.

Dataset

The model was trained using the new-plant-diseases-dataset on Kaggle. The dataset contains approximately 87,000 RGB images of healthy and diseased crop leaves across 38 classes. The dataset was split 80/20 into training and validation sets, and a 33-image test set was created.

Inference Performance

Total AI inference time (preprocessing + model inference): 110 ms