We won the grand prize at the AI hub competition!🥇
We propose a Smart ATM model that can prevent voice phishing face-to-face defraudation damage in ATMs.
-
First, we calculate the risk through personal withdrawal details, loans received from banks, insurance loans, and card loan information.
-
Second, based on the yolov5 algorithm, masks, hands, and faces are detected to determine whether to wear a mask or not, and whether to make a call is determined.
-
Third, based on Efficienet-b4, facial expressions are recognized to determine whether they are embarrassed or anxious.
-
Finally, the risk calculated for each is added to determine whether there is voice phishing.
More detailed description can be found through this link
Our short paper : link
 김영민 |
 곽윤경 |
 여지민 |
 양동재 |
- Ubuntu 18.04.5LTS, Tesla V100-SXM2 32GB
- Facial Recognition : AI hub 한국인 표정인식 데이터셋
- Mask Detection : Kaggle Mask Dataset
- Hand Detection : COCO-Hand Dataset
- My data(Sample Data) : Individual Information(we make)
We use YOLOv5 + EfficientNet-b4. Yolov5 is a real-time object detection model that can quickly detect objects. EfficientNet-b4 is a classification model with high accuracy with a small number of parameters.
$ git clone https://github.com/winston1214/Smart_ATM.git && cd Smart_ATM
$ pip install -r requirements.txt
$ python detect.py --source ${VIDEO_PATH} --weights weights/detection_best.pt --facial-weights-file weights/facial_best.pt --id ${user number}
Object Detection Training
This is the same as yolov5 method.
-
Original yolov5 github : yolov5
-
Train description(Korean ver) : My Blog
-
Train description(Eng ver) : My github readme
Facial Recognition
- Setting dataset(Image) → Crop the face part of a person's image.
- Seeting dataset(Label) → Make a csv file with two columns(image name, label). ※ I set [normal : 0, danger : 1]
- change directory
$ cd facial_recognition - run
$ python facial_train.py --root ${image root} --csv ${label csv} --batch {Batch-size} --epochs {Number of epoch} --lr {learning rate}
- Train loss
- Validation Accuracy
Public._.mp4
if you don't watch a output video, you should click here to watch output video