| import numpy as np |
| import cv2 |
| from dataclasses import dataclass |
| from typing import Optional |
| from src.geometry import FaceGeometry |
|
|
| @dataclass |
| class QualityReport: |
| passed: bool |
| score: float |
| warnings: list[str] |
| blocking: Optional[str] = None |
| blocking_code: Optional[str] = None |
|
|
| def _blocking(code: str, message: str) -> QualityReport: |
| return QualityReport( |
| passed=False, |
| score=0.0, |
| warnings=[], |
| blocking=message, |
| blocking_code=code, |
| ) |
|
|
| def assess_quality(landmarks, img) -> QualityReport: |
| h, w = img.shape[:2] |
| geo = FaceGeometry(landmarks) |
| warnings = [] |
|
|
| gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) |
| brightness = np.mean(gray) |
| if brightness < 40: |
| warnings.append("Image is too dark — try better lighting") |
| elif brightness > 230: |
| warnings.append("Image is overexposed — try softer lighting") |
|
|
| face_width_px = geo.face_width() |
| min_face_px = min(w, h) * 0.15 |
|
|
| if face_width_px < min_face_px: |
| return _blocking( |
| "FACE_TOO_SMALL", |
| "Face too small — move closer to the camera", |
| ) |
| |
| nose_x = geo.nose()[0] |
| left_x = geo.left_eye()[0] |
| right_x = geo.right_eye()[0] |
| eye_mid_x = (left_x + right_x) / 2 |
| eye_dist = abs(left_x - right_x) |
|
|
| yaw_offset = abs(nose_x - eye_mid_x) / eye_dist if eye_dist > 0 else 0 |
| if yaw_offset > 0.35: |
| return _blocking( |
| "FACE_ROTATED", |
| "Please face the camera directly", |
| ) |
| elif yaw_offset > 0.25: |
| warnings.append("Head is turned — results may be less accurate") |
| |
| chin_y = geo.chin()[1] |
| forehead_y = geo.forehead_top()[1] |
| face_h = abs(chin_y - forehead_y) |
| nose_y = geo.nose()[1] |
| expected_y = (chin_y + forehead_y) / 2 |
|
|
| pitch_offset = abs(nose_y - expected_y) / face_h if face_h > 0 else 0 |
| if pitch_offset > 0.25: |
| return _blocking( |
| "FACE_TILTED", |
| "Head is tilted too much — try a straight-on photo", |
| ) |
| elif pitch_offset > 0.15: |
| warnings.append("Head is tilted up or down - try a straight-on photo") |
|
|
| left_cheek_x = geo.left_cheek()[0] |
| right_cheek_x = geo.right_cheek()[0] |
| mid_x = (left_cheek_x + right_cheek_x) / 2 |
| |
| nose_offset = abs(geo.nose()[0] - mid_x) / face_width_px |
| if nose_offset > 0.20: |
| return _blocking( |
| "POOR_ALIGNMENT", |
| "Could not align landmarks — try a clearer, well-lit photo", |
| ) |
| if nose_offset > 0.12: |
| warnings.append("Unusual landmark alignment - lighting or angle may be off") |
| |
| score = max(0.1, 1.0 - len(warnings) * 0.2) |
|
|
| return QualityReport(passed=True, score=round(score, 2), warnings=warnings) |
|
|
