readme

DayBreak-u · Jul 30, 2020 · 5e3cd08 · 5e3cd08
1 parent ba4ccef
commit 5e3cd08
Show file tree

Hide file tree

Showing 18 changed files with 43 additions and 127 deletions.
diff --git a/.DS_Store b/.DS_Store
diff --git a/README.md b/README.md
@@ -1,102 +1,21 @@
-## 本项目基于[chineseocr](https://github.com/chineseocr/chineseocr) 与[psenet](https://github.com/WenmuZhou/PSENet.pytorch)  实现中文自然场景文字检测及识别
-
+## chineseocr_lite 的 onnx 推理
 # 环境
-- pytorch  1.2.0 
 
-- python3
+- python3.6
 
 - linux/macos/windows
----
-- windows环境配置参考热心网友的文章[Python构建快速高效的中文文字识别OCR](https://blog.csdn.net/lly1122334/article/details/104752851) 👍
-- windows环境运行需要手动编译psenet/pse目录下的pse.cpp为DLL，为了方便新增支持python3.5和python3.6的pyd文件，在windows平台运行
-的时候需要根据自己环境的python版本更改对应的pyd文件为pse.pyd即可
-
-## Docker 环境
-- 重写Dockerfile，资源占用更小，可在1C 1G的学生服务器编译成功
-- 容器环境是python3.6 + pytorch-cpu1.2
-- 编译方式：在项目根目录下运行 `docker build -t chineseocr:v1 .`
-- Flask-RESTful API运行方式: `docker run --name ocr_api -p5000:5000 -d <Image ID>`
-  请求方式详见下方
-- web app运行方式：`docker run --name ocr_ui -p8000:8000 -d <Image ID> python3 app.py 8000`
-
-
-## PSENET 编译
-``` Bash
-cd psenet/pse
-rm -rf pse.so 
-make 
-```
-
-# 实现功能
-- [x]  提供轻量的backone检测模型psenet（8.5M）,crnn_lstm_lite(9.5M) 和行文本方向分类网络（1.5M）
-- [x]  任意方向文字检测，识别时判断行文本方向 
-- [x]  crnn\crnn_lite lstm\dense识别（ocr-dense和ocr-lstm是搬运[chineseocr](https://github.com/chineseocr/chineseocr)的）   
-- [x]  支持竖排文本识别  
-- [x]  ncnn 实现 (支持lstm) nihui大佬实现的[crnn_lstm推理](https://github.com/ouyanghuiyu/chineseocr_lite/pull/41) 具体操作详解: [详细记录超轻量中文OCR LSTM模型ncnn实现](https://zhuanlan.zhihu.com/p/113338890?utm_source=qq&utm_medium=social&utm_oi=645149500650557440)
-- [x]  提供竖排文字样例以及字体库（旋转90度的字体）
-- [x]  dbnet ncnn 实现 (感谢 @[zhengqicl](https://github.com/zhengqicl) 的实现)
-
-# 2020.07.02更新
-- 提供dbnet模型，dbnet.onnx(3.7M) dbnet_lite.onnx(1.7M)
-
-
-# 2020.03.16更新
-- psenet ncnn核扩展实现，有效解决粘连文本检测问题，详见[ncnn ocr一条龙](https://github.com/ouyanghuiyu/chineseocr_lite/tree/master/ncnn_project/ocr)
-- nihui大佬实现的[crnn_lstm推理](https://github.com/ouyanghuiyu/chineseocr_lite/pull/41) 具体操作详解: [详细记录超轻量中文OCR LSTM模型ncnn实现](https://zhuanlan.zhihu.com/p/113338890?utm_source=qq&utm_medium=social&utm_oi=645149500650557440)
-
-# 2020.03.12更新
-- 升级crnn_lite_lstm_dw.pth模型crnn_lite_lstm_dw_v2.pth , 精度更高
-
-
-
-## 竖排字体样式：
-  <img width="300" height="200" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/vertical_text_fonts/imgs/test.jpg"/>
-
-## 竖排生成的竖排文本样例：
-  <img width="256" height="32" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/vertical_text_fonts/imgs/00156360.jpg"/>
-  <img width="256" height="32" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/vertical_text_fonts/imgs/00000027.jpg"/>
-  <img width="256" height="32" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/vertical_text_fonts/imgs/00156365.jpg"/>
-  <img width="256" height="32" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/vertical_text_fonts/imgs/00187940.jpg"/>
 
 
 ## web服务启动
 ``` Bash
 cd chineseocr_lite## 进入chineseocr目录
-python app.py 8080 ##8080端口号，可以设置任意端口
+python backend/main.py 
 ```
-## 访问服务
-http://127.0.0.1:8080/ocr
 
 
-## Flask-RESTful API demo
-`cd chineseocr_lite && python flask_app.py`
-```text
-- 请求url: http://ip:port/api/v1/ocr
-- 请求方式：POST
-- 请求参数
-    - ImgString：图片转base64后的字符串， str
-- 返回实例
-{
-    "code": "SUCCESS",
-    "text": "不配图我总觉得不舒服不完整不专业"
-}
-```
 ## 识别结果展示
 
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/test_imgs/5_res.jpg"/>
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/test_imgs/4_res.jpg"/>
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/test_imgs/1_res.jpg"/>
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/test_imgs/2_res.jpg"/>
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/test_imgs/3_res.jpg"/>
-
-
-## ncnn检测识别展示(x86 cpu 单进程)
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/ncnn_project/ocr/res_imgs/res_3.jpg"/>
-<img width="500" height="300" src="https://github.com/ouyanghuiyu/chineseocr_lite/blob/master/ncnn_project/ocr/res_imgs/res_2.jpg"/>
-
-## flask-restful api 内存测试(dbnet)
-200张图片测试稳定在1-1.5G左右的内存
-![](test_imgs/flask-api.png)
+<p align="center"><img src="test_imgs/res.jpg"\></p>
 
 
 ## 参考

diff --git a/backend/webInterface/tr_run.py b/backend/webInterface/tr_run.py
@@ -1,7 +1,7 @@
 
 
 import time
-from model import  text_predict
+from model import  OcrHandle
 import tornado.web
 import tornado.gen
 import tornado.httpserver
@@ -17,7 +17,7 @@
 
 logger = logging.getLogger(log.LOGGER_ROOT_NAME + '.' +__name__)
 
-
+ocrhandle = OcrHandle()
 
 class TrRun(tornado.web.RequestHandler):
     '''
@@ -106,7 +106,7 @@ def post(self):
             new_height = int(img.height / scale + 0.5)
             img = img.resize((new_width, new_height), Image.BICUBIC)
 
-        res = text_predict(img)
+        res = ocrhandle.text_predict(img)
 
         img_detected = img.copy()
         img_draw = ImageDraw.Draw(img_detected)

diff --git a/config.py b/config.py
@@ -5,17 +5,14 @@
 
 
 dbnet_short_size = 960
-crnn_type = "full"
+crnn_lite = True
 
 model_path = os.path.join(father_path, "models/dbnet.onnx")
 
 # crnn相关
-nh = 256
-
-# if crnn_type == "full":
-#     crnn_model_path = os.path.join(father_path, "models/crnn_lstm.onnx")
-# elif crnn_type == "lite":
-crnn_model_path = os.path.join(father_path, "models/crnn_lite_lstm.onnx")
+crnn_model_path = os.path.join(father_path, "models/crnn_lstm.onnx")
+if crnn_lite:
+    crnn_model_path = os.path.join(father_path, "models/crnn_lite_lstm.onnx")
 
 
 from crnn.keys import alphabetChinese as alphabet

diff --git a/model.py b/model.py
@@ -9,52 +9,52 @@
 from dbnet.dbnet_infer import DBNET
 
 
-text_handle = DBNET(model_path, short_size=dbnet_short_size)
 
+class  OcrHandle(object):
+    def __init__(self):
+        self.text_handle = DBNET(model_path, short_size=dbnet_short_size)
+        self.crnn_handle = CRNNHandle(crnn_model_path)
 
-crnn_handle = CRNNHandle(crnn_model_path)
+    def crnnRecWithBox(self,im, boxes_list,score_list):
+        """
+        crnn模型，ocr识别
+        @@model,
+        @@converter,
+        @@im:Array
+        @@text_recs:text box
+        @@ifIm:是否输出box对应的img
 
+        """
+        results = []
+        boxes_list = sorted_boxes(np.array(boxes_list))
+        for index, (box ,score) in enumerate(zip(boxes_list,score_list)):
 
-def crnnRecWithBox(im, boxes_list,score_list):
-    """
-    crnn模型，ocr识别
-    @@model,
-    @@converter,
-    @@im:Array
-    @@text_recs:text box
-    @@ifIm:是否输出box对应的img
+            tmp_box = copy.deepcopy(box)
+            partImg_array = get_rotate_crop_image(im, tmp_box.astype(np.float32))
 
-    """
-    results = []
-    boxes_list = sorted_boxes(np.array(boxes_list))
-    for index, (box ,score) in enumerate(zip(boxes_list,score_list)):
 
-        tmp_box = copy.deepcopy(box)
-        partImg_array = get_rotate_crop_image(im, tmp_box.astype(np.float32))
-
+            partImg = Image.fromarray(partImg_array).convert("RGB")
 
-        partImg = Image.fromarray(partImg_array).convert("RGB")
 
+            partImg_ = partImg.convert('L')
 
-        partImg_ = partImg.convert('L')
+            try:
 
-        try:
+                simPred = self.crnn_handle.predict(partImg_)  ##识别的文本
+            except:
+                continue
 
-            simPred = crnn_handle.predict(partImg_)  ##识别的文本
-        except:
-            continue
-
-        if simPred.strip() != u'':
-            results.append([tmp_box,simPred,score])
+            if simPred.strip() != u'':
+                results.append([tmp_box,simPred,score])
 
-    return results
+        return results
 
 
-def text_predict(img):
-    boxes_list, score_list = text_handle.process(np.asarray(img).astype(np.uint8))
-    result = crnnRecWithBox(np.array(img), boxes_list,score_list)
-    
-    return result
+    def text_predict(self,img):
+        boxes_list, score_list = self.text_handle.process(np.asarray(img).astype(np.uint8))
+        result = self.crnnRecWithBox(np.array(img), boxes_list,score_list)
+
+        return result
 
 
 if __name__ == "__main__":

diff --git a/models/crnn_lite_lstm.onnx b/models/crnn_lite_lstm.onnx
diff --git a/test_imgs/1.jpg b/test_imgs/1.jpg
diff --git a/test_imgs/1_res.jpg b/test_imgs/1_res.jpg
diff --git a/test_imgs/2.jpg b/test_imgs/2.jpg
diff --git a/test_imgs/2_res.jpg b/test_imgs/2_res.jpg
diff --git a/test_imgs/3.jpg b/test_imgs/3.jpg
diff --git a/test_imgs/3_res.jpg b/test_imgs/3_res.jpg
diff --git a/test_imgs/4.jpg b/test_imgs/4.jpg
diff --git a/test_imgs/4_res.jpg b/test_imgs/4_res.jpg
diff --git a/test_imgs/5.jpg b/test_imgs/5.jpg
diff --git a/test_imgs/5_res.jpg b/test_imgs/5_res.jpg
diff --git a/test_imgs/flask-api.png b/test_imgs/flask-api.png
diff --git a/test_imgs/res.jpg b/test_imgs/res.jpg