add Dockerfile and whl package

.gitattributes

@@ -0,0 +1,4 @@
+*.whl filter=lfs diff=lfs merge=lfs -text
+*.a filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,19 @@
+FROM corex:3.2.1-ubuntu20.04-py3.10-slim
+
+WORKDIR /workspace
+
+ENV LD_LIBRARY_PATH=/usr/local/corex/lib:/usr/local/openmpi/lib:
+COPY requirements.txt /workspace/
+COPY whl /workspace/whl
+RUN pip install whl/*.whl -i https://mirrors.aliyun.com/pypi/simple/
+RUN pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/
+RUN pip install numpy==1.23.5 -i https://mirrors.aliyun.com/pypi/simple/
+RUN apt install -y libgl1
+COPY 1.jpg /workspace/
+COPY test.py /workspace/
+COPY app.py /workspace/
+COPY pre_processor.py /workspace/
+COPY run.sh /workspace/
+RUN rm -rf whl
+
+CMD ["./run.sh"]

README.md

@@ -1 +1,26 @@
 # enginex-bi_series-paddleocr
+
+## Build docker image
+```bash
+docker build -t paddleocr:bi .
+```
+
+## 测试
+### 下载模型
+支持 PP-OCRv4及以下版本
+PP-OCRv4模型：
+- det: https://paddleocr.bj.bcebos.com/PP-OCRv4/chinese/ch_PP-OCRv4_det_infer.tar
+- rec: https://paddleocr.bj.bcebos.com/PP-OCRv4/chinese/ch_PP-OCRv4_rec_infer.tar
+- cls: https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar
+
+### 测试
+模型放在`/mnt/models/ocr/`下，运行下面的测试程序，可以识别出示例图片中的文字
+```bash
+python3 test.py
+```
+
+## OCR API Server
+提供了一个`http` API server，可以通过服务的方式加载模型以及进行图片识别，启动命令如下：
+```bash
+python3 app.py
+```

run.sh

@@ -0,0 +1,5 @@
+#! /usr/bin/env bash
+
+unset NVIDIA_VISIBLE_DEVICES
+unset CUDA_VISIBLE_DEVICES
+python3 app.py

test.py

@@ -0,0 +1,84 @@
+from paddleocr import PaddleOCR
+import time
+import cv2
+import numpy as np
+import os
+import logging
+logging.basicConfig(level=logging.INFO)
+
+# ocr = PaddleOCR(
+#     use_doc_orientation_classify=False, 
+#     use_doc_unwarping=False, 
+#     use_textline_orientation=False) # text detection + text recognition
+# ocr = PaddleOCR(use_doc_orientation_classify=True, use_doc_unwarping=True) # text image preprocessing + text detection + textline orientation classification + text recognition
+# ocr = PaddleOCR(use_doc_orientation_classify=False, use_doc_unwarping=False) # text detection + textline orientation classification + text recognition
+ocr = PaddleOCR(
+    det_model_dir="/mnt/models/ocr/ch_PP-OCRv4_det_server_infer",  # https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar
+    rec_model_dir="/mnt/models/ocr/ch_PP-OCRv4_rec_server_infer",  # https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar
+    cls_model_dir="/mnt/models/ocr/ch_ppocr_mobile_v2.0_cls_infer",  # https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar
+    use_doc_orientation_classify=False,
+    use_doc_unwarping=False,
+    use_gpu=True,
+    cpu_threads=32,
+    enable_mkldnn=True,
+    use_textline_orientation=False) # Switch to PP-OCRv5_mobile models
+
+time.sleep(5)
+total_time = 0
+WITH_PREPROCESSING = True
+NOISE_THRESHOLD = int(os.getenv("NOISE_THRESHOLD", "10"))
+for i in range(1):
+    # img_path = f"test{i%5}.pdf"
+    img_path = f"1.jpg"
+    print(f"run for {img_path}")
+    image = cv2.imread(img_path)
+    start = time.time()
+    
+    # height, width = image.shape[:2]
+    # image = cv2.resize(image, (width // 2, height // 2))
+    
+    if WITH_PREPROCESSING:
+            # 计算图像的灰度直方图
+        if len(image.shape) == 3:  # 如果是彩色图像 (具有三个通道)
+            gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+        else:
+            gray = image
+            
+        logging.info(f"cvtColor takes {time.time() - start} s")
+        
+        # 噪声水平估计（简单地使用标准差作为代理）
+        noise_level = np.std(gray)  # 在灰度图像上评估噪声水平
+        logging.info(f"noise_level: {noise_level}")
+        if noise_level > NOISE_THRESHOLD:
+            logging.info("denoising")
+            if len(image.shape) == 3:  # 如果是彩色图像
+                logging.info("denoisin for color")
+                processed_img = cv2.fastNlMeansDenoisingColored(image)
+            else:  # 如果是灰度图像
+                logging.info("denoising for gray")
+                processed_img = cv2.fastNlMeansDenoising(image)
+        else:
+            processed_img = image
+            
+        logging.info(f"Preprocessing takes {time.time() - start} s")
+    else:
+        processed_img = image
+        logging.info("Skip Preprocessing")
+    
+    result = ocr.ocr(image)
+    end = time.time()
+    print(f"{img_path} elapsed: {end - start}")
+    total_time += (end - start)
+    
+print(f"total time: {total_time}")
+for idx in range(len(result)):
+    res = result[idx]
+    print(f"res: {res}")
+    # for line in res:
+    #     print(line)
+
+# result = ocr.predict("docs/images/00006737.jpg")
+# for res in result:
+#     res.print()
+#     res.save_to_img("output")
+#     res.save_to_json("output")