Electrocardiograms are still vastly stored on paper, which restrains the development of data-driven algorithms for their analysis. We propose in this manuscript an efficient method to digitize Electrocardiogram (ECG) signals from paper images, as part of the "Digitization and Classification of ECG Images: The George B. Moody PhysioNet Challenge 2024". A total of 21,799 12-lead recordings from the PTB-XL public data-set were used to generate realistic images. The architecture of our method consists of three main steps. First, we generate ECG paper images with different characteristics such as resolutions, layouts, grid patterns, wrinkles, temperatures, text and noise levels to reproduce real-life challenges. Second, we perform a fully automated ECG image processing pipeline. To do so, a U-Net model is employed for image binarization where only the relevant part of the image is extracted. In addition, a pre-trained YOLO object detection model is fine-tuned to detect the signal box leads. Combining the obtained cleaned ECG image and detected lead boxes, we finally extract 1D ECG signal for each lead present in the ECG paper. By comparing original and estimated ECG signals, our model achieved a signal-to-noise ratio (SNR) score of 0.447 on the hidden validation set.