Identify text/non text

Run the text_non_text.py script passing the name of the directory, which contains the dataset , as a parameter. The script will use a convolutional neural network to classify the images from the dataset between with text e nontext. After the classification, the script will copy the images into 2 subdiretories, DATASET_PROCESSADO/COM_TEXTO and DATASET_PROCESSADO/SEM_TEXTO, according to the respective output of the CNN.

Example

text_non_text.py
dataset
└── images
    └── Img1.jpg
    └── Img2.jpg
    └── Img3.jpg
    └── ...

Following the example of the strucuture of the diretories given above, the execution of the script will be:

python3 text_non_text.py dataset

Don't forget to store the images inside a subdiretory. The parameter passed to the script must be dataset, following the example above, not dataset/images.

Requirements

• Keras 2.5.0
• Numpy 1.19.5
• Tensorflow 2.5.0
• Sklearn 0.24.2
• Matplotlib 3.4.2
• gdow 3.13.0
• tf-nightly 2.6.0

Results

The CNN used is a MobileNet, basead on publication [1]. An accuracy of 92,4% was obtained during the validation step and using a dataset of 1000 images.

Reference

[1] Bai, Xiang, et al. "Text/non-text image classification in the wild with convolutional neural networks." Pattern Recognition 66 (2017): 437-446