Comparative Investigation of Image Feature Extraction Techniques for Text Detection Using UNet, Textsnake, Swin Transformer, and Craft

Text detection is a fundamental task in computer vision with wide-ranging applications including document digitization, autonomous navigation, and assistive technologies. This research presents a comprehensive comparative study of four state-of-the-art deep learning models for image feature extraction in text detection: UNet, TextSnake, Swin Transformer, and CRAFT. Each model embodies a distinct architectural approach segmentation-based, geometry-aware, transformer-based, and character-affinity modeling, respectively. The models are implemented using Python with PyTorch and TensorFlow frameworks, and evaluated on standard benchmark datasets including ICDAR 2021, and the Synth90k dataset. Performance is assessed using objective metrics such as Precision, Recall, F1-Score, Intersection over Union (IoU), Detection Accuracy, Inference Speed (FPS), and Model Size (MB). Experimental results highlight the trade-offs between accuracy, computational efficiency, and generalization capability of each model. Among the models evaluated, CRAFT demonstrates the best overall balance between detection accuracy and robustness, especially in scenarios involving irregular or curved text. These findings provide valuable insights for selecting optimal text detection models based on specific real-world application requirements.