A Modern Tibetan Document Layout Analysis Algorithm Based on DeepLabV3+

In recent years, the field of document layout analysis has faced many technical challenges, especially in terms of multi-scale feature extraction, which has obvious shortcomings. At the same time, it performs poorly in cross-linguistic adaptability, which is particularly prominent in document layout analysis tasks for minority languages such as Tibetan. To tackle these challenges, this study proposes an effective approach based on Convolutional Neural Networks (CNNs), relying on the TibetanAI_TDLA Tibetan document layout dataset to improve the accuracy of element recognition. This paper takes DeepLabV3+ as the foundation and introduces Dilated Convolution (DC), Squeeze and Excitation Attention (SE), and Feature Pyramid Network (FPN) to effectively solve the problem of detail loss in small-scale text and densely arranged elements, significantly improving recognition efficiency. Comparative experiments show that this method achieves a mean Pixel Accuracy (mPA) of 83.1% and a mean Intersection over Union (mIoU) of 84.5%. Compared with other layout analysis models such as U-Net, ARUNet, FCN, SegNet, and DeepLabV3+ by 6.9 %, 9.1 %, 4.1 %, 6.9 % and 2.1 % in mIoU, respectively.