doctra 0.3.2__py3-none-any.whl → 0.4.0__py3-none-any.whl

doctra/parsers/table_chart_extractor.py

@@ -61,22 +61,17 @@ class ChartTablePDFParser:
     ):
         """
         Initialize the ChartTablePDFParser with extraction configuration.
-        
-        Sets up the layout detection engine and optionally the VLM service
-        for structured data extraction.
 
-        :param extract_charts: Whether to extract charts from the document
-        :param extract_tables: Whether to extract tables from the document
-        :param use_vlm: Whether to use VLM for structured data extraction
-        :param vlm_provider: VLM provider to use ("gemini", "openai", "anthropic", or "openrouter")
+        :param extract_charts: Whether to extract charts from the document (default: True)
+        :param extract_tables: Whether to extract tables from the document (default: True)
+        :param use_vlm: Whether to use VLM for structured data extraction (default: False)
+        :param vlm_provider: VLM provider to use ("gemini", "openai", "anthropic", or "openrouter", default: "gemini")
         :param vlm_model: Model name to use (defaults to provider-specific defaults)
-        :param vlm_api_key: API key for VLM provider
-        :param layout_model_name: Layout detection model name
-        :param dpi: DPI for PDF rendering
-        :param min_score: Minimum confidence score for layout detection
-        :raises ValueError: If neither extract_charts nor extract_tables is True
+        :param vlm_api_key: API key for VLM provider (required if use_vlm is True)
+        :param layout_model_name: Layout detection model name (default: "PP-DocLayout_plus-L")
+        :param dpi: DPI for PDF rendering (default: 200)
+        :param min_score: Minimum confidence score for layout detection (default: 0.0)
         """
-        # Validation
         if not extract_charts and not extract_tables:
             raise ValueError("At least one of extract_charts or extract_tables must be True")
 
@@ -98,21 +93,15 @@ class ChartTablePDFParser:
     def parse(self, pdf_path: str, output_base_dir: str = "outputs") -> None:
         """
         Parse a PDF document and extract charts and/or tables.
-        
-        Processes the PDF through layout detection, extracts the specified
-        element types, saves cropped images, and optionally converts them
-        to structured data using VLM.
 
         :param pdf_path: Path to the input PDF file
         :param output_base_dir: Base directory for output files (default: "outputs")
         :return: None
         """
-        # Create output directory structure: outputs/<filename>/structured_parsing/
         pdf_name = Path(pdf_path).stem
         out_dir = os.path.join(output_base_dir, pdf_name, "structured_parsing")
         os.makedirs(out_dir, exist_ok=True)
 
-        # Create subdirectories based on what we're extracting
         charts_dir = None
         tables_dir = None
 
@@ -129,24 +118,20 @@ class ChartTablePDFParser:
         )
         pil_pages = [im for (im, _, _) in render_pdf_to_images(pdf_path, dpi=self.dpi)]
 
-        # Determine which labels to extract
         target_labels = []
         if self.extract_charts:
             target_labels.append("chart")
         if self.extract_tables:
             target_labels.append("table")
 
-        # Count items for progress bars
         chart_count = sum(sum(1 for b in p.boxes if b.label == "chart") for p in pages) if self.extract_charts else 0
         table_count = sum(sum(1 for b in p.boxes if b.label == "table") for p in pages) if self.extract_tables else 0
 
-        # Prepare output content
         if self.use_vlm:
             md_lines: List[str] = ["# Extracted Charts and Tables\n"]
             structured_items: List[Dict[str, Any]] = []
             vlm_items: List[Dict[str, Any]] = []
 
-        # Progress bar descriptions
         charts_desc = "Charts (VLM → table)" if self.use_vlm else "Charts (cropped)"
         tables_desc = "Tables (VLM → table)" if self.use_vlm else "Tables (cropped)"
 
@@ -154,11 +139,9 @@ class ChartTablePDFParser:
         table_counter = 1
 
         with ExitStack() as stack:
-            # Enhanced environment detection
             is_notebook = "ipykernel" in sys.modules or "jupyter" in sys.modules
             is_terminal = hasattr(sys.stdout, 'isatty') and sys.stdout.isatty()
             
-            # Use appropriate progress bars based on environment
             if is_notebook:
                 charts_bar = stack.enter_context(
                     create_notebook_friendly_bar(total=chart_count, desc=charts_desc)) if chart_count else None
@@ -174,23 +157,19 @@ class ChartTablePDFParser:
                 page_num = p.page_index
                 page_img: Image.Image = pil_pages[page_num - 1]
 
-                # Only process selected item types
                 target_items = [box for box in p.boxes if box.label in target_labels]
 
                 if target_items and self.use_vlm:
                     md_lines.append(f"\n## Page {page_num}\n")
 
                 for box in sorted(target_items, key=reading_order_key):
-                    # Handle charts
                     if box.label == "chart" and self.extract_charts:
                         chart_filename = f"chart_{chart_counter:03d}.png"
                         chart_path = os.path.join(charts_dir, chart_filename)
 
-                        # Save image
                         cropped_img = page_img.crop((box.x1, box.y1, box.x2, box.y2))
                         cropped_img.save(chart_path)
 
-                        # Handle VLM processing if enabled
                         if self.use_vlm and self.vlm:
                             rel_path = os.path.join("charts", chart_filename)
                             wrote_table = False
@@ -227,16 +206,13 @@ class ChartTablePDFParser:
                         if charts_bar:
                             charts_bar.update(1)
 
-                    # Handle tables
                     elif box.label == "table" and self.extract_tables:
                         table_filename = f"table_{table_counter:03d}.png"
                         table_path = os.path.join(tables_dir, table_filename)
 
-                        # Save image
                         cropped_img = page_img.crop((box.x1, box.y1, box.x2, box.y2))
                         cropped_img.save(table_path)
 
-                        # Handle VLM processing if enabled
                         if self.use_vlm and self.vlm:
                             rel_path = os.path.join("tables", table_filename)
                             wrote_table = False
@@ -273,19 +249,11 @@ class ChartTablePDFParser:
                         if tables_bar:
                             tables_bar.update(1)
 
-        # Write outputs only if VLM is used
-        md_path = None
         excel_path = None
 
         if self.use_vlm:
-            # Write markdown file
-            md_path = os.path.join(out_dir, "charts.md")
-            with open(md_path, 'w', encoding='utf-8') as f:
-                f.write('\n'.join(md_lines))
 
-            # Write Excel file if we have structured data
             if structured_items:
-                # Determine Excel filename based on extraction target
                 if self.extract_charts and self.extract_tables:
                     excel_filename = "parsed_tables_charts.xlsx"
                 elif self.extract_charts:
@@ -299,23 +267,19 @@ class ChartTablePDFParser:
                 excel_path = os.path.join(out_dir, excel_filename)
                 write_structured_excel(excel_path, structured_items)
                 
-                # Also create HTML version
                 html_filename = excel_filename.replace('.xlsx', '.html')
                 html_path = os.path.join(out_dir, html_filename)
                 write_structured_html(html_path, structured_items)
 
-            # Write VLM items mapping for UI linkage
             if 'vlm_items' in locals() and vlm_items:
                 with open(os.path.join(out_dir, "vlm_items.json"), 'w', encoding='utf-8') as jf:
                     json.dump(vlm_items, jf, ensure_ascii=False, indent=2)
 
-        # Print results
         extraction_types = []
         if self.extract_charts:
             extraction_types.append("charts")
         if self.extract_tables:
             extraction_types.append("tables")
         
-        # Print completion message with output directory
         print(f"✅ Parsing completed successfully!")
         print(f"📁 Output directory: {out_dir}")

doctra/third_party/docres/data/MBD/MBD.py

@@ -0,0 +1,110 @@
+import cv2
+import numpy as np
+import MBD_utils
+import torch
+import torch.nn.functional as F
+
+
+def mask_base_dewarper(image,mask):
+    '''
+    input:
+        image -> ndarray HxWx3 uint8
+        mask -> ndarray HxW uint8
+    return
+        dewarped -> ndarray HxWx3 uint8
+        grid (optional) -> ndarray HxWx2 -1~1
+    '''
+
+    ## get contours
+    # _, contours, hierarchy = cv2.findContours(mask,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_NONE)  ## cv2.__version__ == 3.x
+    contours,hierarchy = cv2.findContours(mask,cv2.RETR_EXTERNAL,method=cv2.CHAIN_APPROX_SIMPLE)  ## cv2.__version__ == 4.x
+
+    ## get biggest contours and four corners based on Douglas-Peucker algorithm
+    four_corners, maxArea, contour= MBD_utils.DP_algorithm(contours)
+    four_corners = MBD_utils.reorder(four_corners)
+
+    ## reserve biggest contours and remove other noisy contours
+    new_mask = np.zeros_like(mask)
+    new_mask = cv2.drawContours(new_mask,[contour],-1,255,cv2.FILLED)
+
+    ## obtain middle points
+    # ratios = [0.25,0.5,0.75]  # ratios = [0.125,0.25,0.375,0.5,0.625,0.75,0.875]
+    ratios = [0.25,0.5,0.75]
+    # ratios = [0.0625,0.125,0.1875,0.25,0.3125,0.375,0.4475,0.5,0.5625,0.625,0.06875,0.75,0.8125,0.875,0.9375]
+    middle = MBD_utils.findMiddle(corners=four_corners,mask=new_mask,points=ratios)
+    
+    ## all points
+    source_points = np.concatenate((four_corners,middle),axis=0) ## all_point = four_corners(topleft,topright,bottom)+top+bottom+left+right
+
+    ## target points
+    h,w = image.shape[:2]
+    padding = 0
+    target_points = [[padding, padding],[w-padding, padding], [padding, h-padding],[w-padding, h-padding]]
+    for ratio in ratios:
+        target_points.append([int((w-2*padding)*ratio)+padding,padding])
+    for ratio in ratios:
+        target_points.append([int((w-2*padding)*ratio)+padding,h-padding])
+    for ratio in ratios:
+        target_points.append([padding,int((h-2*padding)*ratio)+padding])
+    for ratio in ratios:
+        target_points.append([w-padding,int((h-2*padding)*ratio)+padding])
+
+    ## dewarp base on cv2
+    # pts1 = np.float32(source_points)
+    # pts2 = np.float32(target_points)
+    # tps = cv2.createThinPlateSplineShapeTransformer()
+    # matches = []
+    # N = pts1.shape[0]
+    # for i in range(0,N):
+    #     matches.append(cv2.DMatch(i,i,0))
+    # pts1 = pts1.reshape(1,-1,2)
+    # pts2 = pts2.reshape(1,-1,2)
+    # tps.estimateTransformation(pts2,pts1,matches)
+    # dewarped = tps.warpImage(image)
+
+    ## dewarp base on generated grid
+    source_points = source_points.reshape(-1,2)/np.array([image.shape[:2][::-1]]).reshape(1,2)
+    source_points = torch.from_numpy(source_points).float().cuda()
+    source_points = source_points.unsqueeze(0)
+    source_points = (source_points-0.5)*2
+    target_points = np.asarray(target_points).reshape(-1,2)/np.array([image.shape[:2][::-1]]).reshape(1,2)
+    target_points = torch.from_numpy(target_points).float()
+    target_points = (target_points-0.5)*2
+
+    model = MBD_utils.TPSGridGen(target_height=256,target_width=256,target_control_points=target_points)
+    model = model.cuda()
+    grid = model(source_points).view(-1,256,256,2).permute(0,3,1,2)
+    grid = F.interpolate(grid,(h,w),mode='bilinear').permute(0,2,3,1)
+    dewarped = MBD_utils.torch2cvimg(F.grid_sample(MBD_utils.cvimg2torch(image).cuda(),grid))[0]
+    return dewarped,grid[0].cpu().numpy()
+
+def mask_base_cropper(image,mask):
+    '''
+    input:
+        image -> ndarray HxWx3 uint8
+        mask -> ndarray HxW uint8
+    return
+        dewarped -> ndarray HxWx3 uint8
+        grid (optional) -> ndarray HxWx2 -1~1
+    '''
+
+    ## get contours
+    _, contours, hierarchy = cv2.findContours(mask,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_NONE)  ## cv2.__version__ == 3.x
+    # contours,hierarchy = cv2.findContours(mask,cv2.RETR_EXTERNAL,method=cv2.CHAIN_APPROX_SIMPLE)  ## cv2.__version__ == 4.x
+
+    ## get biggest contours and four corners based on Douglas-Peucker algorithm
+    four_corners, maxArea, contour= MBD_utils.DP_algorithm(contours)
+    four_corners = MBD_utils.reorder(four_corners)
+
+    ## reserve biggest contours and remove other noisy contours
+    new_mask = np.zeros_like(mask)
+    new_mask = cv2.drawContours(new_mask,[contour],-1,255,cv2.FILLED)
+
+    ## 最小外接矩形
+    rect = cv2.minAreaRect(contour) # 得到最小外接矩形的（中心(x,y), (宽,高), 旋转角度）
+    box = cv2.boxPoints(rect) # cv2.boxPoints(rect) for OpenCV 3.x 获取最小外接矩形的4个顶点坐标
+    box = np.int0(box)
+    box = box.reshape((4,1,2))    
+
+
+

doctra/third_party/docres/data/MBD/MBD_utils.py

@@ -0,0 +1,291 @@
+import cv2
+import numpy as np
+import copy
+import torch
+import torch
+import itertools
+import torch.nn as nn
+from torch.autograd import Function, Variable
+
+def reorder(myPoints):
+    myPoints = myPoints.reshape((4, 2))
+    myPointsNew = np.zeros((4, 1, 2), dtype=np.int32)
+    add = myPoints.sum(1)
+    myPointsNew[0] = myPoints[np.argmin(add)]
+    myPointsNew[3] =myPoints[np.argmax(add)]
+    diff = np.diff(myPoints, axis=1)
+    myPointsNew[1] =myPoints[np.argmin(diff)]
+    myPointsNew[2] = myPoints[np.argmax(diff)]
+    return myPointsNew
+
+
+def findMiddle(corners,mask,points=[0.25,0.5,0.75]):
+    num_middle_points = len(points)
+    top = [np.array([])]*num_middle_points
+    bottom = [np.array([])]*num_middle_points
+    left = [np.array([])]*num_middle_points
+    right = [np.array([])]*num_middle_points
+
+    center_top = []
+    center_bottom = []
+    center_left = []
+    center_right = []
+
+    center = (int((corners[0][0][1]+corners[3][0][1])/2),int((corners[0][0][0]+corners[3][0][0])/2))
+    for ratio in points:
+
+        center_top.append( (center[0],int(corners[0][0][0]*(1-ratio)+corners[1][0][0]*ratio)) )
+
+        center_bottom.append( (center[0],int(corners[2][0][0]*(1-ratio)+corners[3][0][0]*ratio)) )
+
+        center_left.append( (int(corners[0][0][1]*(1-ratio)+corners[2][0][1]*ratio),center[1]) )
+
+        center_right.append( (int(corners[1][0][1]*(1-ratio)+corners[3][0][1]*ratio),center[1]) )
+
+    for i in range(0,center[0],1):
+        for j in range(num_middle_points):
+            if top[j].size==0:
+                if mask[i,center_top[j][1]]==255:
+                    top[j] = np.asarray([center_top[j][1],i])
+                    top[j] = top[j].reshape(1,2)
+
+    for i in range(mask.shape[0]-1,center[0],-1):
+        for j in range(num_middle_points):
+            if bottom[j].size==0:
+                if mask[i,center_bottom[j][1]]==255:
+                    bottom[j] = np.asarray([center_bottom[j][1],i])
+                    bottom[j] = bottom[j].reshape(1,2)
+
+    for i in range(mask.shape[1]-1,center[1],-1):
+        for j in range(num_middle_points):
+            if right[j].size==0:
+                if mask[center_right[j][0],i]==255:
+                    right[j] = np.asarray([i,center_right[j][0]])
+                    right[j] = right[j].reshape(1,2)
+
+    for i in range(0,center[1]):
+        for j in range(num_middle_points):
+            if left[j].size==0:
+                if mask[center_left[j][0],i]==255:
+                    left[j] = np.asarray([i,center_left[j][0]])
+                    left[j] = left[j].reshape(1,2) 
+
+    return np.asarray(top+bottom+left+right)
+
+def DP_algorithmv1(contours):
+    biggest = np.array([])
+    max_area = 0
+    step = 0.001
+    count = 0
+    # while biggest.size==0:
+    while True:
+        for i in contours:
+            # print(i.shape)
+            area = cv2.contourArea(i)
+            # print(area,cv2.arcLength(i, True))
+            if area > cv2.arcLength(i, True)*10:
+                peri = cv2.arcLength(i, True)
+                approx = cv2.approxPolyDP(i, (0.01+step*count) * peri, True)
+                if area > max_area and len(approx) == 4:
+                    max_area = area
+                    biggest_contours = i
+                    biggest = approx
+                    break
+                    if abs(max_area - cv2.contourArea(biggest))/max_area > 0.3:
+                        biggest = np.array([])
+        count += 1
+        if count > 200:
+            break
+    temp = biggest[0]
+    return biggest,max_area, biggest_contours
+
+def DP_algorithm(contours):
+    biggest = np.array([])
+    max_area = 0
+    step = 0.001
+    count = 0
+
+    ### largest contours
+    for i in contours:
+        area = cv2.contourArea(i)
+        if area > max_area:
+            max_area = area
+            biggest_contours = i
+    peri = cv2.arcLength(biggest_contours, True)
+
+    ### find four corners
+    while True:
+        approx = cv2.approxPolyDP(biggest_contours, (0.01+step*count) * peri, True)
+        if len(approx) == 4:
+            biggest = approx
+            break
+            # if abs(max_area - cv2.contourArea(biggest))/max_area > 0.2:
+            # if abs(max_area - cv2.contourArea(biggest))/max_area > 0.4:
+                # biggest = np.array([])
+        count += 1
+        if count > 200:
+            break
+    return biggest,max_area, biggest_contours
+
+def drawRectangle(img,biggest,color,thickness):
+    cv2.line(img, (biggest[0][0][0], biggest[0][0][1]), (biggest[1][0][0], biggest[1][0][1]), color, thickness)
+    cv2.line(img, (biggest[0][0][0], biggest[0][0][1]), (biggest[2][0][0], biggest[2][0][1]), color, thickness)
+    cv2.line(img, (biggest[3][0][0], biggest[3][0][1]), (biggest[2][0][0], biggest[2][0][1]), color, thickness)
+    cv2.line(img, (biggest[3][0][0], biggest[3][0][1]), (biggest[1][0][0], biggest[1][0][1]), color, thickness)
+    return img
+
+def minAreaRect(contours,img):
+    # biggest = np.array([])
+    max_area = 0
+    for i in contours:
+        area = cv2.contourArea(i)
+        if area > max_area:
+            peri = cv2.arcLength(i, True)
+            rect = cv2.minAreaRect(i)
+            points = cv2.boxPoints(rect)
+            max_area = area
+    return points
+
+def cropRectangle(img,biggest):
+    # print(biggest)
+    w = np.abs(biggest[0][0][0] - biggest[1][0][0])
+    h = np.abs(biggest[0][0][1] - biggest[2][0][1])
+    new_img = np.zeros((w,h,img.shape[-1]),dtype=np.uint8)
+    new_img = img[biggest[0][0][1]:biggest[0][0][1]+h,biggest[0][0][0]:biggest[0][0][0]+w]
+    return new_img
+
+def cvimg2torch(img,min=0,max=1):
+    '''
+    input:
+        im -> ndarray uint8 HxWxC 
+    return
+        tensor -> torch.tensor BxCxHxW 
+    '''
+    if len(img.shape)==2:
+        img = np.expand_dims(img,axis=-1)
+    img = img.astype(float) / 255.0
+    img = img.transpose(2, 0, 1) # NHWC -> NCHW
+    img = np.expand_dims(img, 0)
+    img = torch.from_numpy(img).float()
+    return img
+
+def torch2cvimg(tensor,min=0,max=1):
+    '''
+    input:
+        tensor -> torch.tensor BxCxHxW C can be 1,3
+    return
+        im -> ndarray uint8 HxWxC 
+    '''
+    im_list = []
+    for i in range(tensor.shape[0]):
+        im = tensor.detach().cpu().data.numpy()[i]
+        im = im.transpose(1,2,0)
+        im = np.clip(im,min,max)
+        im = ((im-min)/(max-min)*255).astype(np.uint8)
+        im_list.append(im)
+    return im_list
+
+
+
+class TPSGridGen(nn.Module):
+    def __init__(self, target_height, target_width, target_control_points):
+        '''
+        target_control_points -> torch.tensor  num_pointx2 -1~1
+        source_control_points -> torch.tensor  batch_size x num_point x 2 -1~1
+        return:
+            grid -> batch_size x hw x 2 -1~1
+        '''
+        super(TPSGridGen, self).__init__()
+        assert target_control_points.ndimension() == 2
+        assert target_control_points.size(1) == 2
+        N = target_control_points.size(0)
+        self.num_points = N
+        target_control_points = target_control_points.float()
+
+        # create padded kernel matrix
+        forward_kernel = torch.zeros(N + 3, N + 3)
+        target_control_partial_repr = self.compute_partial_repr(target_control_points, target_control_points)
+        forward_kernel[:N, :N].copy_(target_control_partial_repr)
+        forward_kernel[:N, -3].fill_(1)
+        forward_kernel[-3, :N].fill_(1)
+        forward_kernel[:N, -2:].copy_(target_control_points)
+        forward_kernel[-2:, :N].copy_(target_control_points.transpose(0, 1))
+        # compute inverse matrix
+        inverse_kernel = torch.inverse(forward_kernel)
+
+        # create target cordinate matrix
+        HW = target_height * target_width
+        target_coordinate = list(itertools.product(range(target_height), range(target_width)))
+        target_coordinate = torch.Tensor(target_coordinate) # HW x 2
+        Y, X = target_coordinate.split(1, dim = 1)
+        Y = Y * 2 / (target_height - 1) - 1
+        X = X * 2 / (target_width - 1) - 1
+        target_coordinate = torch.cat([X, Y], dim = 1) # convert from (y, x) to (x, y)
+        target_coordinate_partial_repr = self.compute_partial_repr(target_coordinate.to(target_control_points.device), target_control_points)
+        target_coordinate_repr = torch.cat([
+            target_coordinate_partial_repr, torch.ones(HW, 1), target_coordinate
+        ], dim = 1)
+
+        # register precomputed matrices
+        self.register_buffer('inverse_kernel', inverse_kernel)
+        self.register_buffer('padding_matrix', torch.zeros(3, 2))
+        self.register_buffer('target_coordinate_repr', target_coordinate_repr)
+
+    def forward(self, source_control_points):
+        assert source_control_points.ndimension() == 3
+        assert source_control_points.size(1) == self.num_points
+        assert source_control_points.size(2) == 2
+        batch_size = source_control_points.size(0)
+
+        Y = torch.cat([source_control_points, Variable(self.padding_matrix.expand(batch_size, 3, 2))], 1)
+        mapping_matrix = torch.matmul(Variable(self.inverse_kernel), Y)
+        source_coordinate = torch.matmul(Variable(self.target_coordinate_repr), mapping_matrix)
+        return source_coordinate
+    # phi(x1, x2) = r^2 * log(r), where r = ||x1 - x2||_2
+    def compute_partial_repr(self, input_points, control_points):
+        N = input_points.size(0)
+        M = control_points.size(0)
+        pairwise_diff = input_points.view(N, 1, 2) - control_points.view(1, M, 2)
+        # original implementation, very slow
+        # pairwise_dist = torch.sum(pairwise_diff ** 2, dim = 2) # square of distance
+        pairwise_diff_square = pairwise_diff * pairwise_diff
+        pairwise_dist = pairwise_diff_square[:, :, 0] + pairwise_diff_square[:, :, 1]
+        repr_matrix = 0.5 * pairwise_dist * torch.log(pairwise_dist)
+        # fix numerical error for 0 * log(0), substitute all nan with 0
+        mask = repr_matrix != repr_matrix
+        repr_matrix.masked_fill_(mask, 0)
+        return repr_matrix
+
+
+
+
+ 
+    ### deside wheather further process
+    # point_area = cv2.contourArea(np.concatenate((biggest_angle[0].reshape(1,1,2),middle[0:3],biggest_angle[1].reshape(1,1,2),middle[9:12],biggest_angle[3].reshape(1,1,2),middle[3:6][::-1],biggest_angle[2].reshape(1,1,2),middle[6:9][::-1]),axis=0))
+    #### 最小外接矩形
+    # rect = cv2.minAreaRect(contour) # 得到最小外接矩形的（中心(x,y), (宽,高), 旋转角度）
+    # box = cv2.boxPoints(rect) # cv2.boxPoints(rect) for OpenCV 3.x 获取最小外接矩形的4个顶点坐标
+    # box = np.int0(box)
+    # box = box.reshape((4,1,2))
+    # minrect_area = cv2.contourArea(box)
+    # print(abs(minrect_area-point_area)/point_area)
+    #### 四个角点 IOU
+    # biggest_box = np.concatenate((biggest_angle[0,:,:].reshape(1,1,2),biggest_angle[2,:,:].reshape(1,1,2),biggest_angle[3,:,:].reshape(1,1,2),biggest_angle[1,:,:].reshape(1,1,2)),axis=0)
+    # biggest_mask = np.zeros_like(mask)
+    # # corner_area = cv2.contourArea(biggest_box)
+    # cv2.drawContours(biggest_mask,[biggest_box], -1, color=255, thickness=-1)
+    
+    # smooth = 1e-5
+    # biggest_mask_ = biggest_mask > 50
+    # mask_ = mask > 50
+    # intersection = (biggest_mask_ & mask_).sum()
+    # union = (biggest_mask_ | mask_).sum()
+    # iou = (intersection + smooth) / (union + smooth)
+    # if iou > 0.975:
+    #     skip = True
+    # else:
+    #     skip = False
+    # print(iou)
+    # cv2.imshow('mask',cv2.resize(mask,(512,512)))
+    # cv2.imshow('biggest_mask',cv2.resize(biggest_mask,(512,512)))
+    # cv2.waitKey(0)