Loading repository data…
Neerajj9 / repository
Implemented the YOLO algorithm for scene text detection in keras-tensorflow (No object detection API used) The code can be tweaked to train for a different object detection task using YOLO.
A transparent discovery signal based on current public GitHub metadata.
This score does not audit code, security, maintainers, documentation quality, or suitability. Verify the repository and its current documentation before adoption.
First step towards building an efficient OCR system is to find out the specific text locations. Implemented the YOLO ( You Only Look Once ) algorithm from scratch (no object detection API used) for the specific task of Scene Text Detection in python using keras and tensorflow.
The dataset used is ICDAR competetion dataset available here : Drive Link Train images = 376 Validation images = 115
The Preprocess.py file handles all the necessary preprocessing and saves the data in the form of numpy arrays. First, the images are resized to (512,512) dimensions. Accordingly, the ground truth of the boxes is modified as well. All the images are normalized to a range of [-1 , 1]. The ground truth coordinates are processed to form a matrix of dimensions as ( grid height , grid width , 1 , 5 ).
Necessary changes need to be done in the Preprocess.py file to input the custom data and images, to create the appropriate input and output matrices.
MobileNetv2 architecture is used as a feture extractor. The main reason for choosing MobileNetv2 is the high accuracy and the less number of weights. The fully connected layers of MobileNet are removed. Three Conv layers are added to the last layer of the MobileNet architecture to output a shape of (grid height , grid width , 1 , 5 ). The model weights can be found here : Drive Link
The loss function implemented is the one specified in the YOLO paper. As there is only one class to be predicted, the contribution of class predictions to the loss is eliminated.
The model is trained for 180 epochs in total with a batch size of 4. The learning rate was kept at 0.001 for the first 100 epochs and lowered to 0.0001 for the next 80 epochs.
The Utils.py consists of the functions used to convert the matrix output ( grid height , grid width , 1 , 5 ) of the model to actual predicted bounding boxes. Non max suppression is used to eliminate boxes on the same object as stated in the YOLO paper.
Selected from shared topics, language and repository description—not editorial ratings.
Salvatore-2512 /
This is a Python-based implementation of the YOLOv1 (You Only Look Once) object detection model, presented in a Jupyter Notebook. YOLOv1 treats object detection as a regression problem, predicting bounding boxes and class probabilities directly from full images in a single evaluation.
31/100 health