CPNText-DETR: Dynamic contour initialization and normalized loss for accurate arbitrarily shaped text detection

Key Points

• Accurate text detection plays a crucial role in image processing and computer vision tasks.
• The study implements dynamic contour initialization to better capture varied text shapes and improve accuracy.
• Normalized loss is introduced to enhance model training and refine detection performance.
• These methods may improve applications in optical character recognition and scene text recognition.