cv_uhdm_image-demoireing - Ascend NPU 适配
1. 模型简介
UHDM 图像去摩尔纹模型,基于 ESDNet (Encoder-decoder with Semantic alignment and Dense fusion) 架构,用于去除数字屏幕拍摄时产生的摩尔纹。
2. 昇腾 NPU 适配结果
| 指标 | 值 |
|---|
| Cosine Similarity | 1.000000 |
| Max Abs Error | 0.002218 |
| 平均延迟 | 23.90ms |
| 输出维度 | (1, 3, 64, 64) |
| 推理精度 | float32 |
| 设备 | Ascend 910B4 |
3. 环境要求
| 组件 | 版本 |
|---|
| CANN | 8.5.1 |
| torch_npu | 2.9.0.post1 |
| PyTorch | 2.9.0 |
| Python | 3.11 |
| modelscope | latest |
4. 快速使用
export ASCEND_HOME_PATH=/usr/local/Ascend/cann
export LD_LIBRARY_PATH=/usr/local/Ascend/cann/lib64:$LD_LIBRARY_PATH
python inference.py
python evaluate.py
5. 推理输出证据
Model: iic/cv_uhdm_image-demoireing
Device: npu:0
Output shape: [1, 3, 64, 64]
Output sample: [-0.0885179415345192, 0.3886769711971283, -0.39146584272384644, -0.15365758538246155, -0.45795387029647827]
Latency: 23.90ms
--- CPU vs NPU 精度对比 ---
Cosine Similarity: 1.000000
Max Abs Error: 0.002218
SUCCESS
6. CPU vs NPU 精度对比
| 指标 | CPU (float32) | NPU (float32) | 误差 |
|---|
| Cosine Similarity | 基准 | 1.000000 | 0 |
| Max Abs Error | - | 0.002218 | - |
| 输出维度 | (1, 3, 64, 64) | (1, 3, 64, 64) | 一致 |
7. 模型结构
- Backbone: ESDNet (Encoder-decoder with Semantic alignment)
- Encoder: 3-level encoder with RDB + SAM blocks
- Decoder: 3-level decoder with CSAF fusion
- Input: RGB image (padded to 32-divisible)
- Output: Multi-scale restored images (tuple of 3)