This project aims to develop a deep learning model able to detect a barcode in a given image. The model behind it is Tiny YOLO 3, which a light version of YOLO 3 with less hidden layers in the neural network architecture. This helps significantly reduce the inference time, although its predictive accuracy is lower than YOLO 3 itself. For real time applications, this trade-off can be accepted in most cases.
Here are the 4 steps for this project :
- Implement Tiny YOLO 3 with pretrained weights (80 classes). Using transfer learning, train a model on a set of ~600 barcodes images (90% train / 10% validation).
- Use the model trained for inference on a new image.
- Use pyzbar (python library to read barcodes) to decode the barcode.
- Call OpenfoodFact API to retrieve informations about the product (for food products). The final model can be tested in a streamlit app, by uploading an image and getting the resulting image with a bounding box over the barcode.
-
Install python 3.6+.
-
Install zbar for Mac/Linux:
Linux :
sudo apt-get install libzbar0Mac (make sure brew is installed):
brew install zbarThe zbar DLLs are included with the Windows Python wheels.
-
Clone this repository
git clone <url> -
Install the requirements
pip install -r requirements.txt
The inference result depends on some parameters tuning that can be made in settings.py file, especially for:
- score_threshold
- iou_threshold
These parameters can be changed before starting the app. The app can be started like follows :
- Launch streamlit app :
streamlit run app.py - Upload image and click "Launch barcode detection"
- If the barcode is detected, a bounding box will appear in the image around the barcode.
- If the barcode is decoded, it will show in the screen.
- If the OpenFoodFacts API contains information about the product, it will appear in the product info section.
TBD
Keras implementation of Tiny Yolo v3 : https://github.com/zzh8829/yolov3-tf2