6.2. CV 类算法

本页列出已在爱芯 NPU 平台完成适配的计算机视觉模型。所有模型可通过 AXERA 模型仓库（Hugging Face | ModelScope）下载预编译 axmodel。

运行示例与上板命令见 示例展示。

6.2.1. 目标检测

YOLO 系列

• YOLOv5：检测，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv5-Seg：实例分割，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv7-Face：人脸检测，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv8：检测，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv8-Seg：实例分割，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv8-Pose：姿态估计，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLOv8-Aquarium：水族馆/海洋生物检测，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLO11：检测，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLO11-Seg：实例分割，AX8850N / AX8850 | HuggingFace | ModelScope

• YOLO11-Pose：姿态估计，AX8850N / AX8850 | HuggingFace | ModelScope

• yolo11-obb：旋转框检测，AX8850N / AX8850 / AX8910 | HuggingFace | ModelScope

• yolo26：检测，AX8850N / AX8850 / AX8910 | HuggingFace | ModelScope

• yolo26-seg：实例分割，AX8850N / AX8850 / AX8910 | HuggingFace | ModelScope

• yolo26-pose：姿态估计，AX8850N / AX8850 / AX8910 | HuggingFace | ModelScope

• yolo26-obb：旋转框检测，AX8850N / AX8850 / AX8910 | HuggingFace | ModelScope

开放词汇 / Transformer 检测

• YOLO-World-V2：开放词汇检测，AX8850N / AX8850 | HuggingFace | ModelScope

• RT-DETR：实时端到端检测，AX8850N / AX8850 | HuggingFace | ModelScope

• DEIMv2：端到端检测，AX8850N / AX8850 | HuggingFace | ModelScope

• Deformable-Detr：Deformable DETR，AX8850N / AX8850 | HuggingFace | ModelScope

• OWLViT2：开放词汇视觉检测，AX8850N / AX8850 | HuggingFace | ModelScope

• WeDetect：轻量级开放词汇视觉检测 AX8850N / AX8850 | HuggingFace | ModelScope

行业场景检测

• Helmet-axera：安全帽检测 | HuggingFace | ModelScope

• Gesture-axera：手势识别 | HuggingFace | ModelScope

• E_bike-axera：电动车检测 | HuggingFace | ModelScope

• Package-axera：包裹检测 | HuggingFace | ModelScope

• Plate-axera：车牌检测 | HuggingFace | ModelScope

• Pet-axera：宠物检测 | HuggingFace | ModelScope

• Fall-axera：跌倒检测 | HuggingFace | ModelScope

• Drone-axera：无人机检测 | HuggingFace | ModelScope

• Cow-axera：牛只检测 | HuggingFace | ModelScope

• Person_car-axera：人车检测 | HuggingFace | ModelScope

• QRCode-axera：二维码检测 | HuggingFace | ModelScope

• Bird-Species-Classification：鸟类分类 | HuggingFace | ModelScope

• pp-nsfw_Inspector：内容安全审核 | HuggingFace | ModelScope

6.2.2. 分割

• MobileSAM：轻量 Segment Anything，AX8850N / AX8850 | HuggingFace | ModelScope

• EdgeTAM：边缘端 Track Anything，AX8850N / AX8850 | HuggingFace | ModelScope

• DeepLabv3Plus：语义分割，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.3. 人脸

• Insightface：人脸检测 / 识别 / 关键点，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.4. 深度估计

• Depth-Anything-V2：单目深度估计，AX8850N / AX8850 | HuggingFace | ModelScope

• Depth-Anything-3：单目深度估计 v3，AX8850N / AX8910 | HuggingFace | ModelScope

• IGEV-plusplus：双目立体匹配，AX8850N / AX8850 | HuggingFace | ModelScope

• RAFT-stereo：双目立体匹配，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.5. 图像增强 / 超分 / 修复

• Real-ESRGAN：真实世界超分辨率，AX8850N / AX8850 | HuggingFace | ModelScope

• Real-ESRGAN.axera：Real-ESRGAN 方案包，AX8850N / AX8850 | HuggingFace | ModelScope

• CodeFormer：人脸修复，AX8850N / AX8850 | HuggingFace | ModelScope

• DeOldify：黑白上色，AX8850N / AX8850 | HuggingFace | ModelScope

• SuperResolution：通用超分，AX8850N / AX8850 | HuggingFace | ModelScope

• Z-Image-Turbo：图像增强，AX8850N / AX8850 | HuggingFace | ModelScope

• RIFE.axera：视频插帧，AX8850N / AX8850 | HuggingFace | ModelScope

• RMBG-1.4：背景移除，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.6. 跟踪

• MixFormerV2：单目标跟踪，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.7. CLIP / 图文对齐

• clip：OpenAI CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• FG-CLIP：细粒度 CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• MobileCLIP：轻量 CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• LibCLIP：CLIP 推理库，AX8850N / AX8850 | HuggingFace | ModelScope

• jina-clip-v2：Jina CLIP v2，AX8850N / AX8850 | HuggingFace | ModelScope

• siglip-so400m-patch14-384：SigLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• siglip2-base-patch16-224：SigLIP2，AX8850N / AX8850 | HuggingFace | ModelScope

• cnclip：中文 CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• ViT-L-14-336__axera：ViT-L/14 CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• ViT-L-14-336-CN__axera：ViT-L/14 中文 CLIP，AX8850N / AX8850 | HuggingFace | ModelScope

• dinov3_vits_pretrain_lvd1689m：DINOv3 ViT-S，AX8850N / AX8850 | HuggingFace | ModelScope

• immich：Immich 图像分类模型，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.8. 姿态 / 关键点

• superpoint：特征点检测与描述，AX8850N / AX8910 | HuggingFace | ModelScope

• RTMPose：实时人体姿态估计，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.9. 生成

• lcm-lora-sdv1-5：Stable Diffusion 1.5 + LCM-LoRA，AX8850N / AX8850 | HuggingFace | ModelScope

• LivePortrait：肖像动画驱动，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.10. OCR

• PPOCR_v5：PaddleOCR v5 检测+识别，AX8850N / AX8850 | HuggingFace | ModelScope

• satrn：场景文本识别，AX8850N / AX8850 | HuggingFace | ModelScope

• OCR_RAG-AX8850N：OCR + RAG 方案，AX8850N / AX8850 | HuggingFace | ModelScope

6.2.11. 3D / BEV

• bevformer：BEV 感知，AX8850N / AX8850 | HuggingFace | ModelScope

• centerpoint：3D 目标检测，AX8850N / AX8850 | HuggingFace | ModelScope