 Hi, welcome back. In this part 6 of the series, I will briefly introduce some other features in the photo recognition application, including frame capture, memory dump, and onboard validation. The frame capture mode captures frames from the camera to a microSD card. The memory dump mode dumps every step of the processing pipeline into a microSD card for debugging. The onboard validation mode evaluates the neural network with images stored in the microSD card. So please note that, to use these modes, you will need a microSD card plugged into the STM32 H747 ID scan report, like the picture shown on the right. Next, I will show you how the frame capture works. I will use the binary of the float model from the FPA Vision 1 function pack. This is a demo of how to use a frame capture application. To enter these modes, first press the reset button, then you need to press and hold the blue wake up button during the welcome screen. Then you should see this operating mode screen waiting for your selection. Now, you can start a desired application by pressing the corresponding arrow key of the joystick. This is a joystick. Now let me press the right side to enter frame capture mode. You have to choose a format of the captured files, BMP or IW. Let's press down to select BMP format. If the microSD card is inserted correctly, you should see this live camera capture on the screen. On the top right corner showing the status of the capture ready and on top left corner is a session name. Then you can press the wake up button to capture the frame. You will see the display is frozen and showing capture busy. The captured frame is stored into the microSD card under the folder of the session name. When this process is finished, the capture ready will show up again and you can start the next capture. Next, let's see what is stored in the microSD card. I've pulled out the card, put in the card reader and plugged into my laptop. Open the microSD card. There is a camera capture folder. Open it. You can see the folder name with the session name. This folder includes all the captured frames. There are pictures from the camera frame buffer with a 640 x 480 resolution because we are using the float model and pictures from the input buffer with a 224 x 224 resolution. Now let's see more details about the three different modes. The main purpose of the frame capture application is to enable the data collection. By default, the frame is captured at two stages. One right after the camera acquisition, that's the one we solve with the 640 x 480 resolution. And one after all the preprocessing stages, just before being fed into the input of the neural network, the one we solve with 224 x 224 resolution. But by modifying the application code, the frame can be captured at any stage of the execution chain. For more details, please refer to the user manual 2611. The memory dump application is mainly intended for debug purpose because it's able to dump the memory content for each execution stage into the micro SD card. Operating process is very similar with the frame capture. The difference is after entering this mode, you need to choose the input image source. There are three sources. SD card, the input image is coming from a BMP file stored in the micro SD card. Camera life, the input image is coming from the camera acquisition like we did in the camera capture mode. And test color bar, the input image is also coming from the camera acquisition by the camera is configured in test color bar mode. Same here for more details, please refer to the user manual 2611. Last but not least, the onboard validation. It is very important to note that the onboard validation is different from the validation in the XCube AI tool. The onboard validation uses the raw data without pre-processing, while the XCube AI validation uses the representative dataset. Please refer to the UM 2611 and the XCube AI documentation for more details. Besides, you also need to follow the specific folder structure in the micro SD card for this input images, which is also listed in the UM 2611. And all images must be stored in the BMP format with VGA or QVGA resolution. There is a helper Python script included in the function pack to help you convert the images. You can find it in the directory, utilities, AI resources, food recognition. This is what the onboard validation looks like running on the board. It shows the class name on the top, on the left, you can find the current image being processed, the top one output with the confidence level, and the average loss. And on the right side is the confusion matrix. When all images are processed, a message is displayed on the screen, and then press the wakeup button to update the display with the classification report, like this. The confusion matrix, the list of misclassified files, and the classification report are saved at the root of the micro SD card. This is the figure 16 from the UM 2611, which gives you an overview of validation, capture, and testing modes. The figure shows the processing pipeline with the corresponding outputs. You can find the buffer names, the output file names, and the preview of the output files. Okay, that's all for this video. Thank you for your time. This is the last video about the food recognition application. We might add more videos for other applications in the future, so please stay tuned. At the meantime, you can go to our website, www.sc.com.scm32qai, for the latest information. Thank you and see you then.