# 在应用程序中集成 AI Hub 模型 本节介绍如何在参考应用程序中使用 AI Hub 模型。此示例使用来自 AI Hub 的图像分类模型。 Note 大多数公开的图像分类模型的预处理和后处理都类似，并且可能由 `qtimlpostprocess` 插件支持。 Qualcomm AI Hub 为 Qualcomm 设备提供优化的 AI 模型，允许它们使用 LiteRT（以前称为 Tensorflow Lite）或 Qualcomm® AI Engine Direct 在 CPU、GPU 或 NPU 上运行。 此示例说明如何使用 AI Hub 中的 [inception_v3](https://aihub.qualcomm.com/iot/models/inception_v3) 模型并将该模型集成到 `gst-ai-classification` 参考应用程序中。 | **编号** | **示例程序** | **经过验证的 AI Hub 模型** | **标签文件** | | --- | --- | --- | --- | | 1 | gst-ai-classification | GoogLeNet | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 2 | gst-ai-classification | inception\_v3 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 3 | gst-ai-classification | mobilenet-v2 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 4 | gst-ai-classification | MobileNet-v3-Large | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 5 | gst-ai-classification | ResNet101 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 6 | gst-ai-classification | SqueezeNet-1\_1 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 7 | gst-ai-classification | ResNet18 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 8 | gst-ai-classification | ResNeXt50 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 9 | gst-ai-classification | WideResNet50 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 10 | gst-ai-classification | Shufflenet-v2 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 11 | gst-ai-classification | ResNeXt101 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/classification.json) | | 12 | gst-ai-segmentation | deeplabv3\_plus\_mobilenet | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/deeplabv3_resnet50.json) | | 13 | gst-ai-segmentation | fcn\_resnet50 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/deeplabv3_resnet50.json) | | 14 | gst-ai-segmentation | ffnet\_40s | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/deeplabv3_resnet50.json) | | 15 | gst-ai-segmentation | ffnet\_54s | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/deeplabv3_resnet50.json) | | 16 | gst-ai-segmentation | ffnet\_78s | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/deeplabv3_resnet50.json) | | 17 | gst-ai-object-detection | Yolo-v7 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/yolox.json) | | 18 | gst-ai-object-detection | YOLOv8-Detection | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/yolox.json) | | 19 | GST-AI-superresolution | QuickSRNetLarge | 不适用 | | 20 | GST-AI-superresolution | QuickSRNetMedium | 不适用 | | 21 | GST-AI-superresolution | QuickSRNetSmall | 不适用 | | 22 | GST-AI-superresolution | XLSR | 不适用 | | 23 | gst-ai-monodepth | Midas-V2 | [标签](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/monodepth.json) | | 24 | gst-ai-pose-detection | HRNetPose | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Note 某些模型在 AI Hub 上只有 LiteRT 或 Qualcomm AI Engine Direct SDK 模型。不过，AI Hub 经常更新其中的库以包含更多模型并增强现有模型。 前提条件 1. 下载 [AI Hub 模型](https://aihub.qualcomm.com/iot/models/inception_v3)。 1. 在搜索栏中搜索模型并选择模型。 2. 选择要为其下载模型的 runtime、精度和设备。 - runtime：*TFLite* - 精度：*w8a8* - 设备： - Qualcomm Dragonwing™ RB3 Gen 2: *Qualcomm QCS6490* - Qualcomm Dragonwing™ IQ-9075: *Qualcomm QCS9075* 2. 将与型号匹配的标签文件复制到设备。 > > > 标签文件可通过提供的[链接](https://github.com/quic/sample-apps-for-qualcomm-linux/blob/main/artifacts/json_labels/)下载。 Note 以下部分对通过 AI Hub 下载的模型使用占位符模型名称。将命令中的模型名称更新为通过 AI Hub 下载的相应模型名称。 ## 修改配置 1. 将模型复制到设备。 scp inception_v3-inception-v3-w8a8.tflite root@:/etc/models/ Copy to clipboard 2. 将标签复制到设备。 wget https://raw.githubusercontent.com/quic/sample-apps-for-qualcomm-linux/refs/heads/main/artifacts/json_labels/classification.json Copy to clipboard scp classification.json root@:/etc/labels/ Copy to clipboard 3. 更新 `/etc/configs/config_classification.json` 文件，如下所示。 { "file-path": "/etc/media/video.mp4", "ml-framework": "tflite", "model": "/etc/models/inception_v3-inception-v3-w8a8.tflite", "labels": "/etc/labels/classification.json", "threshold": 40, "runtime": "dsp" } Copy to clipboard ## 运行示例应用程序 1. 运行以下命令，为显示器启用 Weston 服务： export XDG_RUNTIME_DIR=/dev/socket/weston && export WAYLAND_DISPLAY=wayland-1 Copy to clipboard 2. 在命令行参数中指定常量值，运行示例程序： gst-ai-classification --config-file=/etc/configs/config_classification.json Copy to clipboard 3. 使用示例应用程序在用户的设备上运行各种 AI Hub 模型。运行不同的 AI Hub 模型： 1. 下载您想要运行的模型和标签文件。 2. 将模型和标签文件复制到设备。 3. 按照以下步骤运行示例应用程序。 Tab gst-ai-object-detection Tab gst-ai-segmentation Tab gst-ai-superresolution 1. 下载 [Yolox Quantized LiteRT 模型文件](https://aihub.qualcomm.com/iot/models/yolox)（用于目标检测）。 2. 更新 `/etc/configs/config_detection.json` 文件，如下所示。 { "file-path": "/etc/media/video.mp4", "ml-framework": "tflite", "yolo-model-type": "yolox", "model": "/etc/models/yolox_quantized.tflite", "labels": "/etc/labels/yolox.json", "threshold": 40, "runtime": "dsp" } Copy to clipboard Note 根据您获取的模型，更新配置文件中的模型名称。 3. 运行示例应用程序。 Note 确保您已将模型和标签文件复制至设备。 gst-ai-object-detection --config-file=/etc/configs/config_detection.json Copy to clipboard 1. 更新 `/etc/configs/config_segmentation.json` 文件如下： 使用 `deeplabv3_plus_mobilenet_quantized.tflite` 模型文件（用于分割）和 `deeplabv3_resnet50.json` 文件作为示例。 { "file-path": "/etc/media/video.mp4", "ml-framework": "tflite", "model": "/etc/models/deeplabv3_plus_mobilenet_quantized.tflite", "labels": "/etc/labels/deeplabv3_resnet50.json", "runtime": "dsp" } Copy to clipboard 2. 运行示例应用程序。 gst-ai-segmentation --config-file=/etc/configs/config_segmentation.json Copy to clipboard 1. 更新 `/etc/configs/config-superresolution.json` 文件，如下所示。 { "input-file-path": "/etc/media/super.mp4", "model": "/etc/models/quicksrnetsmall_quantized.tflite", "output-file-path":"/etc/media/out.mp4" } Copy to clipboard 2. 运行示例应用： gst-ai-superresolution --config-file=/etc/configs/config-superresolution.json Copy to clipboard Note 确保使用的视频具有 128x128 分辨率，作为模型的输入。 如需了解更多输入参数的详细信息，请参阅应用程序的帮助部分。 Note 在进行分割的情况下，未被识别的对象可能会以白色背景显示。该问题源于标签文件，将在未来的更新中得到解决。 4. 如需进行任何更改，可重新编译示例程序并运行。关于如何使用 eSDK 编译参考应用程序的说明，请参阅 [Qualcomm Intelligent Multimedia 软件开发包 (IM SDK) 快速入门指南](https://docs.qualcomm.com/bundle/publicresource/topics/80-70022-51/introduction.html)。 Last Published: Nov 05, 2025 [Previous Topic 在应用程序中集成定制 AI 模型](https://docs.qualcomm.com/bundle/publicresource/80-70022-15BY/topics/integrate-custom-model.md) [Next Topic 为定制模型添加后处理支持](https://docs.qualcomm.com/bundle/publicresource/80-70022-15BY/topics/add-postprocessing-support-custom-model.md) Source: [https://docs.qualcomm.com/doc/80-70022-15BY/topic/integrate-ai-hub-models.html](https://docs.qualcomm.com/doc/80-70022-15BY/topic/integrate-ai-hub-models.html)