from __future__ import annotations import os import gc import base64 import io import time import shutil import numpy as np import torch import cv2 import ezdxf import gradio as gr from PIL import Image, ImageEnhance from pathlib import Path from typing import List, Union from ultralytics import YOLOWorld, YOLO from ultralytics.engine.results import Results from ultralytics.utils.plotting import save_one_box from transformers import AutoModelForImageSegmentation from torchvision import transforms from scalingtestupdated import calculate_scaling_factor from shapely.geometry import Polygon, Point, MultiPolygon from scipy.interpolate import splprep, splev from scipy.ndimage import gaussian_filter1d from u2net import U2NETP # --------------------- # Create a cache folder for models # --------------------- CACHE_DIR = os.path.join(os.path.dirname(__file__), ".cache") os.makedirs(CACHE_DIR, exist_ok=True) # --------------------- # Custom Exceptions # --------------------- class DrawerNotDetectedError(Exception): """Raised when the drawer cannot be detected in the image""" pass class ReferenceBoxNotDetectedError(Exception): """Raised when the reference box cannot be detected in the image""" pass # --------------------- # Global Model Initialization with caching and print statements # --------------------- print("Loading YOLOWorld model...") start_time = time.time() yolo_model_path = os.path.join(CACHE_DIR, "yolov8x-worldv2.pt") if not os.path.exists(yolo_model_path): print("Caching YOLOWorld model to", yolo_model_path) shutil.copy("yolov8x-worldv2.pt", yolo_model_path) drawer_detector_global = YOLOWorld(yolo_model_path) drawer_detector_global.set_classes(["box"]) print("YOLOWorld model loaded in {:.2f} seconds".format(time.time() - start_time)) print("Loading YOLO reference model...") start_time = time.time() reference_model_path = os.path.join(CACHE_DIR, "last.pt") if not os.path.exists(reference_model_path): print("Caching YOLO reference model to", reference_model_path) shutil.copy("last.pt", reference_model_path) reference_detector_global = YOLO(reference_model_path) print("YOLO reference model loaded in {:.2f} seconds".format(time.time() - start_time)) print("Loading U²-Net model for reference background removal (U2NETP)...") start_time = time.time() u2net_model_path = os.path.join(CACHE_DIR, "u2netp.pth") if not os.path.exists(u2net_model_path): print("Caching U²-Net model to", u2net_model_path) shutil.copy("u2netp.pth", u2net_model_path) u2net_global = U2NETP(3, 1) u2net_global.load_state_dict(torch.load(u2net_model_path, map_location="cpu")) device = "cpu" u2net_global.to(device) u2net_global.eval() print("U²-Net model loaded in {:.2f} seconds".format(time.time() - start_time)) print("Loading BiRefNet model...") start_time = time.time() birefnet_global = AutoModelForImageSegmentation.from_pretrained( "zhengpeng7/BiRefNet", trust_remote_code=True, cache_dir=CACHE_DIR ) torch.set_float32_matmul_precision("high") birefnet_global.to(device) birefnet_global.eval() print("BiRefNet model loaded in {:.2f} seconds".format(time.time() - start_time)) # Define transform for BiRefNet transform_image_global = transforms.Compose([ transforms.Resize((1024, 1024)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) # --------------------- # Model Reload Function (if needed) # --------------------- def unload_and_reload_models(): global drawer_detector_global, reference_detector_global, birefnet_global, u2net_global print("Reloading models...") start_time = time.time() del drawer_detector_global, reference_detector_global, birefnet_global, u2net_global gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() gc.collect() new_drawer_detector = YOLOWorld(os.path.join(CACHE_DIR, "yolov8x-worldv2.pt")) new_drawer_detector.set_classes(["box"]) new_reference_detector = YOLO(os.path.join(CACHE_DIR, "last.pt")) new_birefnet = AutoModelForImageSegmentation.from_pretrained( "zhengpeng7/BiRefNet", trust_remote_code=True, cache_dir=CACHE_DIR ) new_birefnet.to(device) new_birefnet.eval() new_u2net = U2NETP(3, 1) new_u2net.load_state_dict(torch.load(os.path.join(CACHE_DIR, "u2netp.pth"), map_location="cpu")) new_u2net.to(device) new_u2net.eval() drawer_detector_global = new_drawer_detector reference_detector_global = new_reference_detector birefnet_global = new_birefnet u2net_global = new_u2net print("Models reloaded in {:.2f} seconds".format(time.time() - start_time)) # --------------------- # Helper Function: resize_img (defined once) # --------------------- def resize_img(img: np.ndarray, resize_dim): return np.array(Image.fromarray(img).resize(resize_dim)) # --------------------- # Other Helper Functions for Detection & Processing # --------------------- def yolo_detect(image: Union[str, Path, int, Image.Image, list, tuple, np.ndarray, torch.Tensor]) -> np.ndarray: t = time.time() results: List[Results] = drawer_detector_global.predict(image) if not results or len(results) == 0 or len(results[0].boxes) == 0: raise DrawerNotDetectedError("Drawer not detected in the image.") print("Drawer detection completed in {:.2f} seconds".format(time.time() - t)) return save_one_box(results[0].cpu().boxes.xyxy, im=results[0].orig_img, save=False) def detect_reference_square(img: np.ndarray): t = time.time() res = reference_detector_global.predict(img, conf=0.45) if not res or len(res) == 0 or len(res[0].boxes) == 0: raise ReferenceBoxNotDetectedError("Reference box not detected in the image.") print("Reference detection completed in {:.2f} seconds".format(time.time() - t)) return ( save_one_box(res[0].cpu().boxes.xyxy, res[0].orig_img, save=False), res[0].cpu().boxes.xyxy[0] ) # Use U2NETP for reference background removal. def remove_bg_u2netp(image: np.ndarray) -> np.ndarray: t = time.time() image_pil = Image.fromarray(image) transform_u2netp = transforms.Compose([ transforms.Resize((320, 320)), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]), ]) input_tensor = transform_u2netp(image_pil).unsqueeze(0).to("cpu") with torch.no_grad(): outputs = u2net_global(input_tensor) pred = outputs[0] pred = (pred - pred.min()) / (pred.max() - pred.min() + 1e-8) pred_np = pred.squeeze().cpu().numpy() pred_np = cv2.resize(pred_np, (image_pil.width, image_pil.height)) pred_np = (pred_np * 255).astype(np.uint8) print("U2NETP background removal completed in {:.2f} seconds".format(time.time() - t)) return pred_np # Use BiRefNet for main object background removal. def remove_bg(image: np.ndarray) -> np.ndarray: t = time.time() image_pil = Image.fromarray(image) input_images = transform_image_global(image_pil).unsqueeze(0).to("cpu") with torch.no_grad(): preds = birefnet_global(input_images)[-1].sigmoid().cpu() pred = preds[0].squeeze() pred_pil = transforms.ToPILImage()(pred) scale_ratio = 1024 / max(image_pil.size) scaled_size = (int(image_pil.size[0] * scale_ratio), int(image_pil.size[1] * scale_ratio)) result = np.array(pred_pil.resize(scaled_size)) print("BiRefNet background removal completed in {:.2f} seconds".format(time.time() - t)) return result def make_square(img: np.ndarray): height, width = img.shape[:2] max_dim = max(height, width) pad_height = (max_dim - height) // 2 pad_width = (max_dim - width) // 2 pad_height_extra = max_dim - height - 2 * pad_height pad_width_extra = max_dim - width - 2 * pad_width if len(img.shape) == 3: padded = np.pad(img, ((pad_height, pad_height + pad_height_extra), (pad_width, pad_width + pad_width_extra), (0, 0)), mode="edge") else: padded = np.pad(img, ((pad_height, pad_height + pad_height_extra), (pad_width, pad_width + pad_width_extra)), mode="edge") return padded def shrink_bbox(image: np.ndarray, shrink_factor: float): height, width = image.shape[:2] center_x, center_y = width // 2, height // 2 new_width = int(width * shrink_factor) new_height = int(height * shrink_factor) x1 = max(center_x - new_width // 2, 0) y1 = max(center_y - new_height // 2, 0) x2 = min(center_x + new_width // 2, width) y2 = min(center_y + new_height // 2, height) return image[y1:y2, x1:x2] def exclude_scaling_box(image: np.ndarray, bbox: np.ndarray, orig_size: tuple, processed_size: tuple, expansion_factor: float = 1.2) -> np.ndarray: x_min, y_min, x_max, y_max = map(int, bbox) scale_x = processed_size[1] / orig_size[1] scale_y = processed_size[0] / orig_size[0] x_min = int(x_min * scale_x) x_max = int(x_max * scale_x) y_min = int(y_min * scale_y) y_max = int(y_max * scale_y) box_width = x_max - x_min box_height = y_max - y_min expanded_x_min = max(0, int(x_min - (expansion_factor - 1) * box_width / 2)) expanded_x_max = min(image.shape[1], int(x_max + (expansion_factor - 1) * box_width / 2)) expanded_y_min = max(0, int(y_min - (expansion_factor - 1) * box_height / 2)) expanded_y_max = min(image.shape[0], int(y_max + (expansion_factor - 1) * box_height / 2)) image[expanded_y_min:expanded_y_max, expanded_x_min:expanded_x_max] = 0 return image def resample_contour(contour): num_points = 1000 smoothing_factor = 5 spline_degree = 3 if len(contour) < spline_degree + 1: raise ValueError(f"Contour must have at least {spline_degree + 1} points, but has {len(contour)} points.") contour = contour[:, 0, :] tck, _ = splprep([contour[:, 0], contour[:, 1]], s=smoothing_factor) u = np.linspace(0, 1, num_points) resampled_points = splev(u, tck) smoothed_x = gaussian_filter1d(resampled_points[0], sigma=1) smoothed_y = gaussian_filter1d(resampled_points[1], sigma=1) return np.array([smoothed_x, smoothed_y]).T # --------------------- # Add the missing extract_outlines function # --------------------- def extract_outlines(binary_image: np.ndarray) -> (np.ndarray, list): contours, _ = cv2.findContours(binary_image, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE) outline_image = np.zeros_like(binary_image) cv2.drawContours(outline_image, contours, -1, (255), thickness=2) return cv2.bitwise_not(outline_image), contours # --------------------- # Functions for Finger Cut Clearance # --------------------- def union_tool_and_circle(tool_polygon: Polygon, center_inch, circle_diameter=1.0): radius = circle_diameter / 2.0 circle_poly = Point(center_inch).buffer(radius, resolution=64) union_poly = tool_polygon.union(circle_poly) return union_poly def build_tool_polygon(points_inch): return Polygon(points_inch) def polygon_to_exterior_coords(poly: Polygon): if poly.geom_type == "MultiPolygon": biggest = max(poly.geoms, key=lambda g: g.area) poly = biggest if not poly.exterior: return [] return list(poly.exterior.coords) def place_finger_cut_randomly(tool_polygon, points_inch, existing_centers, all_polygons, circle_diameter=1.0, min_gap=0.25, max_attempts=30): import random needed_center_distance = circle_diameter + min_gap radius = circle_diameter / 2.0 for _ in range(max_attempts): idx = random.randint(0, len(points_inch) - 1) cx, cy = points_inch[idx] too_close = False for (ex_x, ex_y) in existing_centers: if np.hypot(cx - ex_x, cy - ex_y) < needed_center_distance: too_close = True break if too_close: continue circle_poly = Point((cx, cy)).buffer(radius, resolution=64) union_poly = tool_polygon.union(circle_poly) overlap_with_others = False too_close_to_others = False for poly in all_polygons: if union_poly.intersects(poly): overlap_with_others = True break if circle_poly.buffer(min_gap).intersects(poly): too_close_to_others = True break if overlap_with_others or too_close_to_others: continue existing_centers.append((cx, cy)) return union_poly, (cx, cy) print("Warning: Could not place a finger cut circle meeting all spacing requirements.") return None, None # --------------------- # DXF Spline and Boundary Functions # --------------------- def save_dxf_spline(inflated_contours, scaling_factor, height, finger_clearance=False): degree = 3 closed = True doc = ezdxf.new(units=0) doc.units = ezdxf.units.IN doc.header["$INSUNITS"] = ezdxf.units.IN msp = doc.modelspace() finger_cut_centers = [] final_polygons_inch = [] for contour in inflated_contours: try: resampled_contour = resample_contour(contour) points_inch = [(x * scaling_factor, (height - y) * scaling_factor) for x, y in resampled_contour] if len(points_inch) < 3: continue if np.linalg.norm(np.array(points_inch[0]) - np.array(points_inch[-1])) > 1e-2: points_inch.append(points_inch[0]) tool_polygon = build_tool_polygon(points_inch) if finger_clearance: union_poly, center = place_finger_cut_randomly(tool_polygon, points_inch, finger_cut_centers, final_polygons_inch, circle_diameter=1.0, min_gap=0.25, max_attempts=30) if union_poly is not None: tool_polygon = union_poly exterior_coords = polygon_to_exterior_coords(tool_polygon) if len(exterior_coords) < 3: continue msp.add_spline(exterior_coords, degree=degree, dxfattribs={"layer": "TOOLS"}) final_polygons_inch.append(tool_polygon) except ValueError as e: print(f"Skipping contour: {e}") return doc, final_polygons_inch def add_rectangular_boundary(doc, polygons_inch, boundary_length, boundary_width, boundary_unit): msp = doc.modelspace() if boundary_unit == "mm": boundary_length_in = boundary_length / 25.4 boundary_width_in = boundary_width / 25.4 else: boundary_length_in = boundary_length boundary_width_in = boundary_width min_x = float("inf") min_y = float("inf") max_x = -float("inf") max_y = -float("inf") for poly in polygons_inch: b = poly.bounds min_x = min(min_x, b[0]) min_y = min(min_y, b[1]) max_x = max(max_x, b[2]) max_y = max(max_y, b[3]) if min_x == float("inf"): print("No tool polygons found, skipping boundary.") return None shape_cx = (min_x + max_x) / 2 shape_cy = (min_y + max_y) / 2 half_w = boundary_width_in / 2.0 half_l = boundary_length_in / 2.0 left = shape_cx - half_w right = shape_cx + half_w bottom = shape_cy - half_l top = shape_cy + half_l rect_coords = [(left, bottom), (right, bottom), (right, top), (left, top), (left, bottom)] from shapely.geometry import Polygon as ShapelyPolygon boundary_polygon = ShapelyPolygon(rect_coords) msp.add_lwpolyline(rect_coords, close=True, dxfattribs={"layer": "BOUNDARY"}) return boundary_polygon def draw_polygons_inch(polygons_inch, image_rgb, scaling_factor, image_height, color=(0,0,255), thickness=2): for poly in polygons_inch: if poly.geom_type == "MultiPolygon": for subpoly in poly.geoms: draw_single_polygon(subpoly, image_rgb, scaling_factor, image_height, color, thickness) else: draw_single_polygon(poly, image_rgb, scaling_factor, image_height, color, thickness) def draw_single_polygon(poly, image_rgb, scaling_factor, image_height, color=(0,0,255), thickness=2): ext = list(poly.exterior.coords) if len(ext) < 3: return pts_px = [] for (x_in, y_in) in ext: px = int(x_in / scaling_factor) py = int(image_height - (y_in / scaling_factor)) pts_px.append([px, py]) pts_px = np.array(pts_px, dtype=np.int32) cv2.polylines(image_rgb, [pts_px], isClosed=True, color=color, thickness=thickness, lineType=cv2.LINE_AA) # --------------------- # Main Predict Function with Finger Cut Clearance, Boundary Box, Annotation and Sharpness Enhancement # --------------------- def predict( image: Union[str, bytes, np.ndarray], offset_inches: float, finger_clearance: str, # "Yes" or "No" add_boundary: str, # "Yes" or "No" boundary_length: float, boundary_width: float, boundary_unit: str, annotation_text: str ): overall_start = time.time() # Convert image to NumPy array if needed. if isinstance(image, str): if os.path.exists(image): image = np.array(Image.open(image).convert("RGB")) else: try: image = np.array(Image.open(io.BytesIO(base64.b64decode(image))).convert("RGB")) except Exception: raise ValueError("Invalid base64 image data") # Apply sharpness enhancement if image is a NumPy array. if isinstance(image, np.ndarray): pil_image = Image.fromarray(image) enhanced_image = ImageEnhance.Sharpness(pil_image).enhance(1.5) image = np.array(enhanced_image) try: t = time.time() drawer_img = yolo_detect(image) print("Drawer detection completed in {:.2f} seconds".format(time.time() - t)) t = time.time() shrunked_img = make_square(shrink_bbox(drawer_img, 0.90)) del drawer_img gc.collect() print("Image shrinking completed in {:.2f} seconds".format(time.time() - t)) except DrawerNotDetectedError: raise DrawerNotDetectedError("Drawer not detected! Please take another picture with a drawer.") try: t = time.time() reference_obj_img, scaling_box_coords = detect_reference_square(shrunked_img) print("Reference square detection completed in {:.2f} seconds".format(time.time() - t)) except ReferenceBoxNotDetectedError: raise ReferenceBoxNotDetectedError("Reference box not detected! Please take another picture with a reference box.") t = time.time() reference_obj_img = make_square(reference_obj_img) reference_square_mask = remove_bg_u2netp(reference_obj_img) print("Reference image processing completed in {:.2f} seconds".format(time.time() - t)) t = time.time() try: cv2.imwrite("mask.jpg", cv2.cvtColor(reference_obj_img, cv2.COLOR_RGB2GRAY)) scaling_factor = calculate_scaling_factor( reference_image_path="./Reference_ScalingBox.jpg", target_image=reference_square_mask, feature_detector="ORB", ) except ZeroDivisionError: scaling_factor = None print("Error calculating scaling factor: Division by zero") except Exception as e: scaling_factor = None print(f"Error calculating scaling factor: {e}") if scaling_factor is None or scaling_factor == 0: scaling_factor = 1.0 print("Using default scaling factor of 1.0 due to calculation error") gc.collect() print("Scaling factor determined: {}".format(scaling_factor)) t = time.time() orig_size = shrunked_img.shape[:2] objects_mask = remove_bg(shrunked_img) processed_size = objects_mask.shape[:2] objects_mask = exclude_scaling_box(objects_mask, scaling_box_coords, orig_size, processed_size, expansion_factor=1.2) objects_mask = resize_img(objects_mask, (shrunked_img.shape[1], shrunked_img.shape[0])) del scaling_box_coords gc.collect() print("Object masking completed in {:.2f} seconds".format(time.time() - t)) t = time.time() offset_pixels = (offset_inches / scaling_factor) * 2 + 1 if scaling_factor != 0 else 1 dilated_mask = cv2.dilate(objects_mask, np.ones((int(offset_pixels), int(offset_pixels)), np.uint8)) del objects_mask gc.collect() print("Mask dilation completed in {:.2f} seconds".format(time.time() - t)) # Save the dilated mask for debugging if needed. Image.fromarray(dilated_mask).save("./outputs/scaled_mask_new.jpg") # --- Extract outlines (only used for DXF generation) --- t = time.time() outlines, contours = extract_outlines(dilated_mask) print("Outline extraction completed in {:.2f} seconds".format(time.time() - t)) # Instead of drawing the original contours, we now prepare a clean copy of the shrunk image for drawing new contours. output_img = shrunked_img.copy() del shrunked_img gc.collect() # --- Generate DXF using the extracted contours and apply finger clearance --- t = time.time() use_finger_clearance = True if finger_clearance.lower() == "yes" else False doc, final_polygons_inch = save_dxf_spline(contours, scaling_factor, processed_size[0], finger_clearance=use_finger_clearance) del contours gc.collect() print("DXF generation completed in {:.2f} seconds".format(time.time() - t)) boundary_polygon = None if add_boundary.lower() == "yes": boundary_polygon = add_rectangular_boundary(doc, final_polygons_inch, boundary_length, boundary_width, boundary_unit) if boundary_polygon is not None: final_polygons_inch.append(boundary_polygon) # --- Annotation Text Placement (Centered horizontally) --- min_x = float("inf") min_y = float("inf") max_x = -float("inf") max_y = -float("inf") for poly in final_polygons_inch: b = poly.bounds if b[0] < min_x: min_x = b[0] if b[1] < min_y: min_y = b[1] if b[2] > max_x: max_x = b[2] if b[3] > max_y: max_y = b[3] margin = 0.5 text_x = (min_x + max_x) / 2 text_y = min_y - margin msp = doc.modelspace() if annotation_text.strip(): text_entity = msp.add_text( annotation_text.strip(), dxfattribs={ "height": 0.25, "layer": "ANNOTATION" } ) text_entity.dxf.insert = (text_x, text_y) dxf_filepath = os.path.join("./outputs", "out.dxf") doc.saveas(dxf_filepath) # --- Draw only the new contours (final_polygons_inch) on the clean output image --- draw_polygons_inch(final_polygons_inch, output_img, scaling_factor, processed_size[0], color=(0,0,255), thickness=2) # Also prepare an "Outlines" image based on a blank canvas for clarity. new_outlines = np.ones_like(output_img) * 255 draw_polygons_inch(final_polygons_inch, new_outlines, scaling_factor, processed_size[0], color=(0,0,255), thickness=2) if annotation_text.strip(): text_px = int(text_x / scaling_factor) text_py = int(processed_size[0] - (text_y / scaling_factor)) cv2.putText(output_img, annotation_text.strip(), (text_px, text_py), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,0,255), 2, cv2.LINE_AA) cv2.putText(new_outlines, annotation_text.strip(), (text_px, text_py), cv2.FONT_HERSHEY_SIMPLEX, 1, (0,0,255), 2, cv2.LINE_AA) outlines_color = cv2.cvtColor(new_outlines, cv2.COLOR_BGR2RGB) print("Total prediction time: {:.2f} seconds".format(time.time() - overall_start)) return ( cv2.cvtColor(output_img, cv2.COLOR_BGR2RGB), outlines_color, dxf_filepath, dilated_mask, str(scaling_factor) ) # --------------------- # Gradio Interface # --------------------- if __name__ == "__main__": os.makedirs("./outputs", exist_ok=True) def gradio_predict(img, offset, finger_clearance, add_boundary, boundary_length, boundary_width, boundary_unit, annotation_text): return predict(img, offset, finger_clearance, add_boundary, boundary_length, boundary_width, boundary_unit, annotation_text) iface = gr.Interface( fn=gradio_predict, inputs=[ gr.Image(label="Input Image"), gr.Number(label="Offset value for Mask (inches)", value=0.075), gr.Dropdown(label="Add Finger Clearance?", choices=["Yes", "No"], value="No"), gr.Dropdown(label="Add Rectangular Boundary?", choices=["Yes", "No"], value="No"), gr.Number(label="Boundary Length", value=300.0, precision=2), gr.Number(label="Boundary Width", value=200.0, precision=2), gr.Dropdown(label="Boundary Unit", choices=["mm", "inches"], value="mm"), gr.Textbox(label="Annotation (max 20 chars)", max_length=20, placeholder="Type up to 20 characters") ], outputs=[ gr.Image(label="Output Image"), gr.Image(label="Outlines of Objects"), gr.File(label="DXF file"), gr.Image(label="Mask"), gr.Textbox(label="Scaling Factor (inches/pixel)") ], examples=[ ["./Test20.jpg", 0.075, "No", "No", 300.0, 200.0, "mm", "MyTool"], ["./Test21.jpg", 0.075, "Yes", "Yes", 300.0, 200.0, "mm", "Tool2"] ] ) iface.launch(share=True)