import cv2 import numpy as np import json import gradio as gr import os import xml.etree.ElementTree as ET # ---------------- Helper functions ---------------- def get_rotated_rect_corners(x, y, w, h, rotation_deg): rot_rad = np.deg2rad(rotation_deg) cos_r, sin_r = np.cos(rot_rad), np.sin(rot_rad) R = np.array([[cos_r, -sin_r], [sin_r, cos_r]]) cx, cy = x + w/2, y + h/2 local_corners = np.array([[-w/2,-h/2],[w/2,-h/2],[w/2,h/2],[-w/2,h/2]]) rotated_corners = np.dot(local_corners, R.T) return (rotated_corners + np.array([cx,cy])).astype(np.float32) def preprocess_gray_clahe(img): gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8,8)) return clahe.apply(gray) def detect_and_match(img1_gray, img2_gray, method="SIFT", ratio_thresh=0.78): if method=="SIFT": detector=cv2.SIFT_create(nfeatures=5000); matcher=cv2.BFMatcher(cv2.NORM_L2) elif method=="ORB": detector=cv2.ORB_create(5000); matcher=cv2.BFMatcher(cv2.NORM_HAMMING) elif method=="BRISK": detector=cv2.BRISK_create(); matcher=cv2.BFMatcher(cv2.NORM_HAMMING) elif method=="KAZE": detector=cv2.KAZE_create(); matcher=cv2.BFMatcher(cv2.NORM_L2) elif method=="AKAZE": detector=cv2.AKAZE_create(); matcher=cv2.BFMatcher(cv2.NORM_HAMMING) else: return None,None,[] kp1, des1 = detector.detectAndCompute(img1_gray,None) kp2, des2 = detector.detectAndCompute(img2_gray,None) if des1 is None or des2 is None: return None,None,[] raw_matches = matcher.knnMatch(des1,des2,k=2) good = [m for m,n in raw_matches if m.distance < ratio_thresh*n.distance] return kp1, kp2, good def parse_xml_points(xml_file): tree = ET.parse(xml_file) root = tree.getroot() points=[] for pt_type in ["TopLeft","TopRight","BottomLeft","BottomRight"]: elem=root.find(f".//point[@type='{pt_type}']") points.append([float(elem.get("x")), float(elem.get("y"))]) return np.array(points,dtype=np.float32).reshape(-1,2) # ---------------- Padding Helper ---------------- def pad_to_size(img, target_h, target_w): h, w = img.shape[:2] top_pad = 0 left_pad = 0 bottom_pad = target_h - h right_pad = target_w - w canvas = np.ones((target_h, target_w,3), dtype=np.uint8)*255 canvas[top_pad:top_pad+h, left_pad:left_pad+w] = img return canvas # ---------------- Resize feature-match to original reference size ---------------- def match_img_to_reference(match_img, ref_h, ref_w): h, w = match_img.shape[:2] scale = min(ref_w/w, ref_h/h) new_w, new_h = int(w*scale), int(h*scale) resized = cv2.resize(match_img, (new_w,new_h)) padded = pad_to_size(resized, ref_h, ref_w) return padded # ---------------- Main Function ---------------- def homography_all_detectors(flat_file, persp_file, json_file, xml_file): flat_img = cv2.imread(flat_file) persp_img = cv2.imread(persp_file) mockup = json.load(open(json_file.name)) roi_data = mockup["printAreas"][0]["position"] roi_x, roi_y = roi_data["x"], roi_data["y"] roi_w, roi_h = mockup["printAreas"][0]["width"], mockup["printAreas"][0]["height"] roi_rot_deg = mockup["printAreas"][0]["rotation"] flat_gray = preprocess_gray_clahe(flat_img) persp_gray = preprocess_gray_clahe(persp_img) xml_points = parse_xml_points(xml_file.name) methods = ["SIFT","ORB","BRISK","KAZE","AKAZE"] gallery_paths = [] download_files = [] for method in methods: kp1,kp2,good_matches = detect_and_match(flat_gray,persp_gray,method) if kp1 is None or kp2 is None or len(good_matches)<4: continue match_img = cv2.drawMatches(flat_img,kp1,persp_img,kp2,good_matches,None,flags=2) src_pts = np.float32([kp1[m.queryIdx].pt for m in good_matches]).reshape(-1,1,2) dst_pts = np.float32([kp2[m.trainIdx].pt for m in good_matches]).reshape(-1,1,2) H,_ = cv2.findHomography(src_pts,dst_pts,cv2.RANSAC,5.0) if H is None: continue roi_corners_flat = get_rotated_rect_corners(roi_x,roi_y,roi_w,roi_h,roi_rot_deg) roi_corners_persp = cv2.perspectiveTransform(roi_corners_flat.reshape(-1,1,2),H).reshape(-1,2) persp_roi = persp_img.copy() cv2.polylines(persp_roi,[roi_corners_persp.astype(int)],True,(0,255,0),2) for px,py in roi_corners_persp: cv2.circle(persp_roi,(int(px),int(py)),5,(255,0,0),-1) xml_gt_img = persp_img.copy() xml_mapped = cv2.perspectiveTransform(xml_points.reshape(-1,1,2),H).reshape(-1,2) for px,py in xml_mapped: cv2.circle(xml_gt_img,(int(px),int(py)),5,(0,0,255),-1) # Convert to RGB flat_rgb = cv2.cvtColor(flat_img,cv2.COLOR_BGR2RGB) persp_rgb = cv2.cvtColor(persp_img,cv2.COLOR_BGR2RGB) roi_rgb = cv2.cvtColor(persp_roi,cv2.COLOR_BGR2RGB) xml_rgb = cv2.cvtColor(xml_gt_img,cv2.COLOR_BGR2RGB) # Resize feature-match image to match original flat/perspective match_rgb = match_img_to_reference(cv2.cvtColor(match_img, cv2.COLOR_BGR2RGB), flat_rgb.shape[0], flat_rgb.shape[1]) # Determine max height and width for grid (all images now same) max_h = max(flat_rgb.shape[0], match_rgb.shape[0], roi_rgb.shape[0], xml_rgb.shape[0]) max_w = max(flat_rgb.shape[1], match_rgb.shape[1], roi_rgb.shape[1], xml_rgb.shape[1]) flat_pad = pad_to_size(flat_rgb, max_h, max_w) roi_pad = pad_to_size(roi_rgb, max_h, max_w) xml_pad = pad_to_size(xml_rgb, max_h, max_w) # Merge 2x2 grid top = np.hstack([flat_pad, match_rgb]) bottom = np.hstack([roi_pad, xml_pad]) combined_grid = np.vstack([top, bottom]) base_name = os.path.splitext(os.path.basename(persp_file))[0] file_name = f"{base_name}_{method.lower()}.png" cv2.imwrite(file_name, cv2.cvtColor(combined_grid,cv2.COLOR_RGB2BGR)) gallery_paths.append(file_name) download_files.append(file_name) while len(download_files)<5: download_files.append(None) return gallery_paths, download_files[0], download_files[1], download_files[2], download_files[3], download_files[4] iface = gr.Interface( fn=homography_all_detectors, inputs=[ gr.Image(label="Upload Flat Image",type="filepath"), gr.Image(label="Upload Perspective Image",type="filepath"), gr.File(label="Upload mockup.json",file_types=[".json"]), gr.File(label="Upload XML file",file_types=[".xml"]) ], outputs=[ gr.Gallery(label="Results per Detector",show_label=True), gr.File(label="Download SIFT Result"), gr.File(label="Download ORB Result"), gr.File(label="Download BRISK Result"), gr.File(label="Download KAZE Result"), gr.File(label="Download AKAZE Result") ], title="Homography ROI Projection with Feature Matching & XML GT", description="Flat + Perspective images with mockup.json & XML. Feature-match aligned with original images using white padding." ) iface.launch()