Multi Model CNN Based Gas Meter Characters Recognition

In the realm of computer vision, the identification and extraction of text from photographs of natural scenes is a very difficult task. Convolutional neural networks (CNNs) have been demonstrated to be highly successful in recognizing characters and words from photographs due to their exceptional ability to recognize the structural patterns within them. This makes CNNs one of the most suitable approaches for solving recognition problems, such as text recognition in natural scene images. In this study, we aim to recognize numerical texts from images and employ three models for this task: CNN models, a combination of CNN-LSTM models, and a combination of CNN-GRU models. The gas meter images comprising the dataset for this study were captured by our team using multiple phones at different times. Our goal is to recognize the counter part of the gas meter digits, which is why we first preprocess the images and prepare the data for detection. We perform the detection using the EAST algorithm. On the gas meter dataset, after the detection stage, we apply data augmentation, normalize image cropping, and perform image equalization. Our results show that the accuracy achieved by the CNN, CNN-LSTM, and CNN-GRU models in recognizing numerical texts from images is 72.9%, 96.6%, and 97.63%, respectively. These findings suggest that the CNN-LSTM and CNN-GRU models are highly effective in recognizing numerical texts in images, with the CNN-GRU model exhibiting the highest accuracy. Overall, these results demonstrate the potential of using deep learning models for recognizing numerical texts in images, particularly the combination of CNN and Gated recurrent unit (GRU) models.