Automated document image classification using custom CNN and pre-trained models, achieving 97% accuracy by leveraging document layout and structure for classification