Hi all. I would like to share my issue in order to get some help.
The task is a binary image classifier including 2 classes {Live, Spoof}.
1. At First, I have trained CNN network as a Face anti spoofing model, with 3 backbones including {Resnet50, MobilenetV2, VGG19} on celebA spoof dataset. among these backbones, VGG19 (superior) outperforms the other backbones (trained with Binary cross entropy).
2. in the next step, I fine-tuned VGG19 checkpoint on my own dataset which is more real faces compared to CelebA dataset. my own dataset train images have 1965 live samples and 6673 spoof samples. I have uploaded the distribution of my own dataset.
since the number of spoof images is 3.5 times larger than live images, I used both weighted loss and focal loss. but the model is still biased to spoof images.
the problem is that my model is biased to spoof classes. it performs weakly on Live class.
I have uploaded 3 plots including {confusion_matrix, AUC, and reliability diagaram} to show the result of fine-tuning on my test set.
in the confusion matrix accuracy of :
- live class is %63
- spoof class is %95
AUC is %79.
My question is, which alternative models can learn better features and discriminate live features from spoof features?
How Can I improve the accuracy of my model on Live images from %63 to > %90 (based on my accuracy on the confusion matrix)
o improve your face anti-spoofing model's accuracy on live images:
Thanks, Usha(@Usha manjari sikharam) for your answer
Please consider that:
1. I don't have access to my subjects to take extra live photos from them so I can't oversample live samples of my own dataset.
2. about offline augmentation, I have limited choices because in face anti-spoofing we can not use any kind of augmentation, it will impact model performance.
3. about using other losses, I used triplet loss and changed my dataloader according to that loss but the result was terrible.
4. I optimized my model hyperparameters and those metric images that I have uploaded, obtained from optimized hyperparameters.
What kind of post-processing techniques do you mean for refining model prediction? Could you mention those techniques?
Hi, instead of fine-tuning the entire VGG19 model, consider freezing some of the initial layers responsible for low-level feature extraction and fine-tuning only the later layers. This approach can help the model focus specifically on the features that differentiate live faces from spoof faces in your dataset.
Thanks Atefeh, but consider that fine-tuning means that I froze some layers and those results are exactly based on freezing initial layers of VGG19.
Thanks Amitesh, currently I have only access to the images of employees. the business requirements are based on the image scenario.
- Badges
- Science topic
- Similar topics
- Social Science
- Media
- Journalism