PP OCRv5 Server Rec

🚀 PP-OCRv5_server_rec

PP-OCRv5_server_rec 是 PaddleOCR 團隊開發的最新一代文本行識別模型 PP-OCRv5_rec 中的一員。它旨在通過單一模型高效、準確地支持簡體中文、繁體中文、英文和日文這四種主要語言的識別，以及手寫、豎排文本、拼音和生僻字等複雜文本場景的識別。

🚀 快速開始

📦 安裝指南

安裝 PaddlePaddle

請參考以下命令，使用 pip 安裝 PaddlePaddle：

# 適用於 CUDA11.8
python -m pip install paddlepaddle-gpu==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu118/

# 適用於 CUDA12.6
python -m pip install paddlepaddle-gpu==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cu126/

# 適用於 CPU
python -m pip install paddlepaddle==3.0.0 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/

PaddlePaddle 安裝的詳細信息，請參考 PaddlePaddle 官方網站。

安裝 PaddleOCR

從 PyPI 安裝最新版本的 PaddleOCR 推理包：

python -m pip install paddleocr

💻 使用示例

基礎用法

from paddleocr import TextRecognition
model = TextRecognition(model_name="PP-OCRv5_server_rec")
output = model.predict(input="2PZfbirjfxA88695lRmgk.jpeg", batch_size=1)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/res.json")

運行後，得到的結果如下：

{'res': {'input_path': '/root/.paddlex/predict_input/2PZfbirjfxA88695lRmgk.jpeg', 'page_index': None, 'rec_text': 'day as a reminder of the', 'rec_score': 0.9534505009651184}}

可視化圖像如下：

使用命令和參數說明的詳細信息，請參考 文檔。

🔧 技術細節

模型精度指標

注意：如果一行中的任何字符（包括標點符號）識別錯誤，則整行標記為錯誤。這確保了在實際應用中具有更高的準確性。

模型管道使用

單個模型的能力是有限的，但由多個模型組成的管道可以提供更強的能力來解決現實場景中的難題。

PP-OCRv5

通用 OCR 管道用於通過從圖像中提取文本信息並以字符串格式輸出，來解決文本識別任務。管道中有 5 個模塊：

• 文檔圖像方向分類模塊（可選）
• 文本圖像去畸變模塊（可選）
• 文本行方向分類模塊（可選）
• 文本檢測模塊
• 文本識別模塊

運行以下命令可快速體驗 OCR 管道：

paddleocr ocr -i https://cdn-uploads.huggingface.co/production/uploads/681c1ecd9539bdde5ae1733c/3ul2Rq4Sk5Cn-l69D695U.png \
    --text_recognition_model_name PP-OCRv5_server_rec \
    --use_doc_orientation_classify False \
    --use_doc_unwarping False \
    --use_textline_orientation True \
    --save_path ./output \
    --device gpu:0 

結果將打印到終端：

{'res': {'input_path': '/root/.paddlex/predict_input/3ul2Rq4Sk5Cn-l69D695U.png', 'page_index': None, 'model_settings': {'use_doc_preprocessor': True, 'use_textline_orientation': True}, 'doc_preprocessor_res': {'input_path': None, 'page_index': None, 'model_settings': {'use_doc_orientation_classify': False, 'use_doc_unwarping': False}, 'angle': -1}, 'dt_polys': array([[[ 352,  105],
        ...,
        [ 352,  128]],

       ...,

       [[ 632, 1431],
        ...,
        [ 632, 1447]]], dtype=int16), 'text_det_params': {'limit_side_len': 64, 'limit_type': 'min', 'thresh': 0.3, 'max_side_limit': 4000, 'box_thresh': 0.6, 'unclip_ratio': 1.5}, 'text_type': 'general', 'textline_orientation_angles': array([0, ..., 0]), 'text_rec_score_thresh': 0.0, 'rec_texts': ['Algorithms for the Markov Entropy Decomposition', 'Andrew J. Ferris and David Poulin', 'Département de Physique, Université de Sherbrooke, Québec, JlK 2R1, Canada', '(Dated: October 31, 2018)', 'The Markov entropy decomposition (MED) is a recently-proposed, cluster-based simulation method for fi-', 'nite temperature quantum systems with arbitrary geometry. In this paper, we detail numerical algorithms for', 'performing the required steps of the MED, principally solving a minimization problem with a preconditioned', 'arXiv:1212.1442v1 [cond-mat.stat-mech] 6Dec 2012', "Newton's algorithm, as well as how to extract global susceptibilities and thermal responses. We demonstrate", 'the power of the method with the spin-1/2 XXZ model on the 2D square lattice, including the extraction of', 'critical points and details of each phase. Although the method shares some qualitative similarities with exact-', 'diagonalization, we show the MED is both more accurate and significantly more flexible.', 'PACS numbers: 05.10.−a,02.50.Ng, 03.67.−a,74.40.Kb', 'I. INTRODUCTION', 'This approximation becomes exact in the case of a 1D quan', 'tum (or classical) Markov chain [10], and leads to an expo-', 'Although the equations governing quantum many-body', 'nential reduction of cost for exact entropy calculations when', 'systems are simple to write down, finding solutions for the', 'the global density matrix is a higher-dimensional Markov net-', 'majority of systems remains incredibly difficult. Modern', 'work state [12, 13].', 'physics finds itself in need of new tools to compute the emer-', 'The second approximation used in the MED approach is', 'gent behavior of large, many-body systems.', 'related to the N-representibility problem. Given a set of lo-', 'There has been a great variety of tools developed to tackle', 'cal but overlapping reduced density matrices {pi}, it is a very', 'many-body problems, but in general, large 2D and 3D quan-', 'challenging problem to determine if there exists a global den-', 'tum systems remain hard to deal with. Most systems are', 'sity operator which is positive semi-definite and whose partial', 'thought to be non-integrable, so exact analytic solutions are', 'trace agrees with each ρi. This problem is QMA-hard (the', 'not usually expected. Direct numerical diagonalization can be', 'quantum analogue of NP) [14, 15], and is hopelessly diffi-', 'performed for relatively small systems — however the emer-', 'cult to enforce. Thus, the second approximation employed', 'gent behavior of a system in the thermodynamic limit may be', 'involves ignoring global consistency with a positive opera-', 'difficult to extract, especially in systems with large correlation', 'tor, while requiring local consistency on any overlapping re-', 'lengths. Monte Carlo approaches are technically exact (up to', 'gions between the ρi. At the zero-temperature limit, the MED', 'sampling error), but suffer from the so-called sign problem', 'approach becomes analogous to the variational nth-order re-', 'for fermionic, frustrated, or dynamical problems. Thus we are', 'duced density matrix approach, where positivity is enforced', 'limited to search for clever approximations to solve the ma-', 'on all reduced density matrices of size n [16–18].', 'jority of many-body problems.', 'The MED approach is an extremely flexible cluster method.', 'Over the past century, hundreds of such approximations', 'applicable to both translationally invariant systems of any di-', 'have been proposed, and we will mention just a few notable', 'mension in the thermodynamic limit, as well as finite systems', 'examples applicable to quantum lattice models. Mean-field', 'or systems without translational invariance (e.g. disordered', 'theory is simple and frequently arrives at the correct quali-', 'lattices, or harmonically trapped atoms in optical lattices).', 'tative description, but often fails when correlations are im-', 'The free energy given by MED is guaranteed to lower bound', 'portant. Density-matrix renormalisation group (DMRG) [1]', 'the true free energy, which in turn lower-bounds the ground', 'is efficient and extremely accurate at solving 1D problems,', 'state energy — thus providing a natural complement to varia-', 'but the computational cost grows exponentially with system', 'tional approaches which upper-bound the ground state energy.', 'size in two- or higher-dimensions [2, 3]. Related tensor-', 'The ability to provide a rigorous ground-state energy window', 'network techniques designed for 2D systems are still in their', 'is a powerful validation tool, creating a very compelling rea-', 'infancy [4–6].  Series-expansion methods [7] can be success-', 'son to use this approach.', 'ful, but may diverge or otherwise converge slowly, obscuring', 'In this paper we paper we present a pedagogical introduc-', 'the state in certain regimes. There exist a variety of cluster-', 'tion to MED, including numerical implementation issues and', 'based techniques, such as dynamical-mean-field theory [8]', 'applications to 2D quantum lattice models in the thermody-', 'and density-matrix embedding [9]', 'namiclimit.InSec.II.wegiveabriefderiyationofthe', 'Here we discuss the so-called Markov entropy decompo-', 'Markov entropy decomposition. Section III outlines a robust', 'sition (MED), recently proposed by Poulin & Hastings [10]', 'numerical strategy for optimizing the clusters that make up', '(and analogous to a slightly earlier classical algorithm [11]).', 'the decomposition. In Sec. IV we show how we can extend', 'This is a self-consistent cluster method for finite temperature', 'these algorithms to extract non-trivial information, such as', 'systems that takes advantage of an approximation of the (von', 'specific heat and susceptibilities. We present an application of', 'Neumann) entropy. In [10], it was shown that the entropy', 'the method to the spin-1/2 XXZ model on a 2D square lattice', 'per site can be rigorously upper bounded using only local in-', 'in Sec. V, describing how to characterize the phase diagram', 'formation — a local, reduced density matrix on N sites, say.', 'and determine critical points, before concluding in Sec. VI.'], 'rec_scores': array([0.99273533, ..., 0.95753616]), 'rec_polys': array([[[ 352,  105],
        ...,
        [ 352,  128]],

       ...,

       [[ 632, 1431],
        ...,
        [ 632, 1447]]], dtype=int16), 'rec_boxes': array([[ 352, ...,  128],
       ...,
       [ 632, ..., 1447]], dtype=int16)}}

如果指定了 save_path，可視化結果將保存在 save_path 下。可視化輸出如下：

命令行方法適用於快速體驗。對於項目集成，也只需要幾行代碼：

from paddleocr import PaddleOCR  

ocr = PaddleOCR(
    text_recognition_model_name="PP-OCRv5_server_rec",
    use_doc_orientation_classify=False, # 使用 use_doc_orientation_classify 啟用/禁用文檔方向分類模型
    use_doc_unwarping=False, # 使用 use_doc_unwarping 啟用/禁用文檔去畸變模塊
    use_textline_orientation=True, # 使用 use_textline_orientation 啟用/禁用文本行方向分類模型
    device="gpu:0", # 使用 device 指定 GPU 進行模型推理
)
result = ocr.predict("https://cdn-uploads.huggingface.co/production/uploads/681c1ecd9539bdde5ae1733c/3ul2Rq4Sk5Cn-l69D695U.png")  
for res in result:  
    res.print()  
    res.save_to_img("output")  
    res.save_to_json("output")

管道中使用的默認模型是 PP-OCRv5_server_rec，你也可以通過參數 text_recognition_model_dir 使用本地模型文件。使用命令和參數說明的詳細信息，請參考 文檔。

PP-StructureV3

佈局分析是一種從文檔圖像中提取結構化信息的技術。PP-StructureV3 包括以下六個模塊：

• 佈局檢測模塊
• 通用 OCR 子線
• 文檔圖像預處理子線（可選）
• 表格識別子線（可選）
• 印章識別子線（可選）
• 公式識別子線（可選）

運行以下命令可快速體驗 PP-StructureV3 管道：

paddleocr pp_structurev3 -i https://cdn-uploads.huggingface.co/production/uploads/681c1ecd9539bdde5ae1733c/mG4tnwfrvECoFMu-S9mxo.png \
    --text_recognition_model_name PP-OCRv5_server_rec \
    --use_doc_orientation_classify False \
    --use_doc_unwarping False \
    --use_textline_orientation False \
    --device gpu:0

結果將打印到終端：

如果指定了 save_path，可視化結果將保存在 save_path 下。可視化輸出如下：

只需幾行代碼即可體驗管道的推理。以 PP-StructureV3 管道為例：

from paddleocr import PPStructureV3

pipeline = PPStructureV3(
    text_recognition_model_name="PP-OCRv5_server_rec",
    use_doc_orientation_classify=False, # 使用 use_doc_orientation_classify 啟用/禁用文檔方向分類模型
    use_doc_unwarping=False,    # 使用 use_doc_unwarping 啟用/禁用文檔去畸變模塊
    use_textline_orientation=False, # 使用 use_textline_orientation 啟用/禁用文本行方向分類模型
    device="gpu:0", # 使用 device 指定 GPU 進行模型推理
    )
output = pipeline.predict("./pp_structure_v3_demo.png")
for res in output:
    res.print() # 打印結構化預測輸出
    res.save_to_json(save_path="output") ## 以 JSON 格式保存當前圖像的結構化結果
    res.save_to_markdown(save_path="output") ## 以 Markdown 格式保存當前圖像的結果

管道中使用的默認模型是 PP-OCRv5_server_rec，你也可以通過參數 text_recognition_model_dir 使用本地模型文件。使用命令和參數說明的詳細信息，請參考 文檔。

📚 詳細文檔

• PaddleOCR 倉庫
• PaddleOCR 文檔

📄 許可證

本項目採用 Apache-2.0 許可證。