Cardiovascular diseases are a major world health challenge, so accurate and timely diagnosis is called for. Manual electrocardiogram (ECG) interpretation is time-consuming and expert-based, so automation is needed. Although there are studies where machine learning (ML) and deep learning (DL) are used for the classification of ECG, these studies depend on digital structured data, neglecting pitfalls in paper-based ECG records. This work presents a new pipeline that turns ECG images into 1D signals for automated cardiac abnormality classification. The system uses advanced preprocessing steps such as grayscale conversion, removal of grid lines, extraction of the signal based on contours, and segmentation into 12-lead based signals. The feature extraction and optimization are done by Principal Component Analysis (PCA) and t-SNE. Various ML models (SVM, KNN, Logistic Regression, XGBoost) and DL models (MLP, LSTM) are tested by hyperparameter optimization. The Voting Classifier recognizes 92.11% cases of cardiac abnormalities in digitized ECG images.