AA-CLIP | 零样本异常检测从训练到部署

OpenCV学堂

发布于 2026-04-02 20:23:58

1650

AA-CLIP介绍

AA-CLIP- 两阶段异常感知CLIP，是最新的零样本异常检测模型，支持文本与图像输入编码，特征比对实现工业与医疗图像的异常分类与分割。官方地址如下：

https://github.com/Mwxinnn/AA-CLIP

AA-CLIP训练

说实话官方github上面的代码写的挺好的，基本没有什么坑，可以直接运行训练，我用我自己的数据集，直接训练，采用小样本模式，八个样本训练，命令行如下：

训练日志如下：

INFO:__main__:args: {'model_name': 'ViT-L-14-336', 'img_size': 518, 'surgery_until_layer': 20, 'relu': False, 'dataset': 'MVTec', 'training_mode': 'few_shot', 'shot': 8, 'text_batch_size': 16, 'image_batch_size': 2, 'text_epoch': 5, 'image_epoch': 20, 'text_lr': 1e-05, 'image_lr': 0.0005, 'criterion': ['dice_loss', 'focal_loss'], 'seed': 111, 'save_path': 'ckpt/baseline', 'text_norm_weight': 0.1, 'text_adapt_weight': 0.1, 'image_adapt_weight': 0.1, 'text_adapt_until': 3, 'image_adapt_until': 6}
INFO:root:Loading pretrained ViT-L-14-336 from OpenAI.
INFO:root:Resizing position embedding grid-size from (24, 24) to (37, 37)
INFO:root:Loading pretrained ViT-L-14-336 from OpenAI.
INFO:root:Resizing position embedding grid-size from (24, 24) to (37, 37)
INFO:__main__:loading dataset ...
INFO:__main__:loading image adaptation dataset ...
INFO:__main__:training text epoch 0:
INFO:__main__:loss: 1.4636036157608032
INFO:__main__:training text epoch 1:
INFO:__main__:loss: 1.219490071137746
INFO:__main__:training text epoch 2:
INFO:__main__:loss: 1.2085641225179036
INFO:__main__:training text epoch 3:
INFO:__main__:loss: 1.2048145135243733
INFO:__main__:training text epoch 4:
INFO:__main__:loss: 1.1965291897455852
INFO:__main__:training image epoch 0:
INFO:__main__:loss: 3.6282341423488798
INFO:__main__:training image epoch 1:
INFO:__main__:loss: 2.6655399345216297
INFO:__main__:training image epoch 2:
INFO:__main__:loss: 2.4324947538829984
INFO:__main__:training image epoch 3:
INFO:__main__:loss: 2.1065715295927867
INFO:__main__:training image epoch 4:
INFO:__main__:loss: 1.9446473377091544
INFO:__main__:training image epoch 5:
INFO:__main__:loss: 1.8010127643744152
INFO:__main__:training image epoch 6:
INFO:__main__:loss: 1.7255745700427465
INFO:__main__:training image epoch 7:
INFO:__main__:loss: 1.737629310006187
INFO:__main__:training image epoch 8:
INFO:__main__:loss: 1.6446210316249303
INFO:__main__:training image epoch 9:
INFO:__main__:loss: 1.5891813351994468
INFO:__main__:training image epoch 10:
INFO:__main__:loss: 1.556674992754346
INFO:__main__:training image epoch 11:
INFO:__main__:loss: 1.432623612029212
INFO:__main__:training image epoch 12:
INFO:__main__:loss: 1.5059531927108765
INFO:__main__:training image epoch 13:
INFO:__main__:loss: 1.4267199337482452
INFO:__main__:training image epoch 14:
INFO:__main__:loss: 1.3787336079847246
INFO:__main__:training image epoch 15:
INFO:__main__:loss: 1.2949471984590804
INFO:__main__:training image epoch 16:
INFO:__main__:loss: 1.3198628212724413
INFO:__main__:training image epoch 17:
INFO:__main__:loss: 1.2334311718032474
INFO:__main__:training image epoch 18:
INFO:__main__:loss: 1.2762719293435414
INFO:__main__:training image epoch 19:
INFO:__main__:loss: 1.1888043440523601

推理测试

训练完以后，官方的代码test.py只支持数据集评估跟eval metrix计算，为了推理任意单张图像，我只能自己一通修改代码，改了一天改好

了。

可以推理单张图像书anomalymap跟mask，运行结果如下：

显示代码如下：

gray_mask = (pixel_preds * 255).astype(np.uint8)
print("prediction score -->> ", preds_image[0])
colored_img = cv.applyColorMap(gray_mask, cv.COLORMAP_JET)
ch, cw, cc = colored_img.shape
print("colored_img-->> ", colored_img.shape)
src = cv.imread(image_file, cv.IMREAD_COLOR)
resized = cv.resize(src, (ch, cw))
anomaly_mask = cv.addWeighted(resized, 0.8, colored_img, 0.6, 0)
cv.imshow("input image", resized)
cv.imshow("anomaly map", colored_img)
cv.imshow("anomaly mask", anomaly_mask)
cv.waitKey(0)
cv.destroyAllWindows()

学习OpenVINO2025

本文参与腾讯云自媒体同步曝光计划，分享自微信公众号。

原始发表：2025-08-20，如有侵权请联系 cloudcommunity@tencent.com 删除

部署