Insightface

基于Insightface的人脸识别Pipeline，使用其buffalo_l系列模型中的人脸检测+定点模型，人脸属性模型，人脸特征提取模型实现人脸识别，支持650N系列平台的1v1人脸比对和1vN人脸识别。

支持芯片:

• AX650N

支持硬件

• M4N-Dock(爱芯派Pro)
• M.2 Accelerator card

原始模型请参考

• Insightface Github

性能对比

Models	input size	Latency (ms)	CMM Usage (MB)
det_10g	640x640	6.947	14.0
genderage	96x96	0.295	0.62
w600k_r50	112x112	3.993	44.0

模型转换

• 模型转换工具链Pulsar2
• 转换文档[TODO]

环境准备

• NPU Python API: pyaxengine

安装需要的python库 pip install -r requirements.txt

运行（算力卡版本）

(insightface) axera@dell:~/insightface/ax$ python insightface_pipeline.py -h
[INFO] Available providers:  ['AXCLRTExecutionProvider']
usage: insightface_pipeline.py [-h] [--model_path MODEL_PATH] [--type TYPE] [--gallery_path GALLERY_PATH] [--query_path QUERY_PATH] [--draw]

Face Recognition Pipeline Example

options:
  -h, --help            show this help message and exit
  --model_path, -m MODEL_PATH
                        Path to the model directory
  --type, -t TYPE       Type of operation: 1: 1v1 compare, 2: 1vN recognize
  --gallery_path, -g GALLERY_PATH
                        Path to the gallery image for image file
  --query_path, -q QUERY_PATH
                        Path to the query image
  --draw, -d            Whether to draw results on the image


(insightface) axera@dell:~/insightface/ax$ python insightface_pipeline.py -g ./gallery_imgs/ -d -t 1
[INFO] Available providers:  ['AXCLRTExecutionProvider']
ModelRouter ./models/buffalo_l/1k3d68.axmodel
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 5.0-patch1-dirty 5cf5ab99-dirty
Applied providers: AXCLRTExecutionProvider, with options: None
model not recognized: ./models/buffalo_l/1k3d68.axmodel
ModelRouter ./models/buffalo_l/2d106det.axmodel
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 5.0-patch1-dirty 5cf5ab99-dirty
Applied providers: AXCLRTExecutionProvider, with options: None
model not recognized: ./models/buffalo_l/2d106det.axmodel
ModelRouter ./models/buffalo_l/det_10g.axmodel
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 5.0-patch1-dirty 5cf5ab99-dirty
Applied providers: AXCLRTExecutionProvider, with options: None
find model: ./models/buffalo_l/det_10g.axmodel detection [1, 640, 640, 3] 0.0 1.0
ModelRouter ./models/buffalo_l/genderage.axmodel
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 5.0-patch1-dirty 5cf5ab99-dirty
Applied providers: AXCLRTExecutionProvider, with options: None
find model: ./models/buffalo_l/genderage.axmodel genderage [1, 96, 96, 3] 0.0 1.0
ModelRouter ./models/buffalo_l/w600k_r50.axmodel
[INFO] Using provider: AXCLRTExecutionProvider
[INFO] SOC Name: AX650N
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Compiler version: 5.0-patch1-dirty 5cf5ab99-dirty
Applied providers: AXCLRTExecutionProvider, with options: None
find model: ./models/buffalo_l/w600k_r50.axmodel recognition [1, 112, 112, 3] 0.0 1.0
set det-size: (640, 640)
warning: det_size is already set in detection model, ignore
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2/2 [00:00<00:00, 12.49it/s]
特征库构建完成，包含以下人员： ['wade', 'kobe']
请输入查询图像路径 (输入 'exit' 退出):
./query/wade2.jpg
识别结果: wade, 相似度(0-1): 0.5860
结果已保存到 ./output/wade2.jpg
请输入查询图像路径 (输入 'exit' 退出):
./query/ssd_horse.jpg
识别结果: Unknown, 相似度(0-1): 0.0135
结果已保存到 ./output/ssd_horse.jpg
请输入查询图像路径 (输入 'exit' 退出):
exit

保存结果在 ./output 目录下：