Ancient book text detection algorithm based on a differentiable binarization network for intangible cultural heritage protection

With the advancement of modernization, the protection of intangible cultural heritage is facing unprecedented challenges. Due to the aging material and long age of ancient texts, they face problems such as weathering and blurred handwriting. Consequently, this study proposes an ancient book text detection algorithm based on differential binarization network, aiming to improve the recognition accuracy of handwritten characters in ancient books and achieve the protection and inheritance of cultural heritage ancient books. This algorithm uses multiple iterations of projection optimization to improve text segmentation performance, and constructs a high-quality annotated dataset based on accurate segmentation results. Meanwhile, an adaptive scale fusion module is introduced to improve the network structure, and a dual detection head parallel structure is adopted to process single character and whole line text. The results show that the method achieves Pixel Accuracy values of 96.50, 97.89, and 98.76% for single character, whole line text, and background. The highest rates of missed and false detection are merely 0.020 and 0.032, respectively. The algorithm has accuracy, recall, and F1 score of 97.12, 96.83, and 97.48%. To sum up, the detection algorithm studied has effectively improved the accuracy and efficiency of ancient text detection, providing strong technical support for the digital protection of intangible cultural heritage.