Tune photo quality thresholds

Author: ShengkaiXu

evaluation_function/evaluation_test.py

@@ -153,25 +153,28 @@ def test_image_quality_rejects_low_single_component_score(self):
         )
         self.assertLessEqual(
             quality["sharpness_score_100"],
-            quality["quality_min_component_score"],
+            quality["quality_min_sharpness_score"],
         )
         self.assertFalse(quality["quality_pass"])
 
     def test_image_quality_soft_warning_does_not_request_retake(self):
         image = np.full((120, 160), 130, dtype=np.uint8)
-        for y in range(0, 120, 20):
-            image[y:y + 10, :] = 170
-        for x in range(0, 160, 20):
-            image[:, x:x + 10] = 90
-        image = cv2.GaussianBlur(image, (3, 3), 0)
+        for y in range(0, 120, 8):
+            image[y:y + 4, :] = 155
+        for x in range(0, 160, 8):
+            image[:, x:x + 4] = 105
+        image = cv2.GaussianBlur(image, (5, 5), 0)
         image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
 
         quality = compute_image_quality_metrics(image)
         advice = " ".join(quality["quality_advice"]).lower()
 
         self.assertTrue(quality["quality_pass"])
-        self.assertGreater(quality["sharpness_score_100"], quality["quality_min_component_score"])
-        self.assertLessEqual(quality["sharpness_score_100"], 50)
+        self.assertGreater(quality["sharpness_score_100"], quality["quality_min_sharpness_score"])
+        self.assertLessEqual(
+            quality["sharpness_score_100"],
+            quality["quality_warning_sharpness_score"],
+        )
         self.assertIn("a little blurry", advice)
         self.assertNotIn("retake", advice)
 
@@ -180,7 +183,7 @@ def test_quality_advice_soft_warnings_cover_all_components(self):
             brightness_mean=50.0,
             brightness_component=0.35,
             contrast_component=0.35,
-            sharpness_component=0.35,
+            sharpness_component=0.20,
             noise_component=0.35,
         )).lower()
 

evaluation_function/yolo_pipeline.py

@@ -71,6 +71,9 @@
 # not rejected while obviously dark/blurry/noisy ones are sent back for retake.
 QUALITY_ACCEPT_SCORE: int = int(os.environ.get("QUALITY_ACCEPT_SCORE", "40"))
 QUALITY_MIN_COMPONENT_SCORE: int = int(os.environ.get("QUALITY_MIN_COMPONENT_SCORE", "20"))
+QUALITY_MIN_SHARPNESS_SCORE: int = int(os.environ.get("QUALITY_MIN_SHARPNESS_SCORE", "10"))
+QUALITY_WARNING_COMPONENT_SCORE: int = int(os.environ.get("QUALITY_WARNING_COMPONENT_SCORE", "50"))
+QUALITY_WARNING_SHARPNESS_SCORE: int = int(os.environ.get("QUALITY_WARNING_SHARPNESS_SCORE", "25"))
 QUALITY_BRIGHTNESS_USABLE_MIN: float = 35.0
 QUALITY_BRIGHTNESS_IDEAL_MIN: float = 60.0
 QUALITY_BRIGHTNESS_IDEAL_MAX: float = 210.0
@@ -790,22 +793,22 @@ def _quality_advice(
 
     if brightness_score <= QUALITY_MIN_COMPONENT_SCORE:
         advice.append(brightness_fail)
-    elif brightness_score <= 50:
+    elif brightness_score <= QUALITY_WARNING_COMPONENT_SCORE:
         advice.append(brightness_warn)
 
     if contrast_score <= QUALITY_MIN_COMPONENT_SCORE:
         advice.append("The parts do not stand out clearly enough for reliable checking. Please retake the photo with a plain background and fewer shadows.")
-    elif contrast_score <= 50:
+    elif contrast_score <= QUALITY_WARNING_COMPONENT_SCORE:
         advice.append("The parts do not stand out very clearly. Use a plain background and avoid shadows next time.")
 
-    if sharpness_score <= QUALITY_MIN_COMPONENT_SCORE:
+    if sharpness_score <= QUALITY_MIN_SHARPNESS_SCORE:
         advice.append("The photo is too blurry for reliable checking. Please retake it after holding the camera still and refocusing.")
-    elif sharpness_score <= 50:
+    elif sharpness_score <= QUALITY_WARNING_SHARPNESS_SCORE:
         advice.append("The photo is a little blurry. For better results, hold the camera still and refocus next time.")
 
     if noise_score <= QUALITY_MIN_COMPONENT_SCORE:
         advice.append("The photo is too noisy or grainy for reliable checking. Please retake it with better lighting and avoid digital zoom.")
-    elif noise_score <= 50:
+    elif noise_score <= QUALITY_WARNING_COMPONENT_SCORE:
         advice.append("The photo looks a little noisy or grainy. Use better lighting and avoid digital zoom next time.")
 
     if not advice:
@@ -848,24 +851,21 @@ def compute_image_quality_metrics(img_bgr: np.ndarray) -> Dict[str, Any]:
     )
     quality_score = int(round(100.0 * (
         0.25 * brightness_component
-        + 0.20 * contrast_component
-        + 0.40 * sharpness_component
-        + 0.15 * noise_component
+        + 0.25 * contrast_component
+        + 0.25 * sharpness_component
+        + 0.25 * noise_component
     )))
     if contrast_component <= 0.0 and sharpness_component <= 0.0:
         quality_score = min(quality_score, QUALITY_ACCEPT_SCORE - 1)
     brightness_score_100 = _score100(brightness_component)
     contrast_score_100 = _score100(contrast_component)
     sharpness_score_100 = _score100(sharpness_component)
     noise_score_100 = _score100(noise_component)
-    component_pass = all(
-        score > QUALITY_MIN_COMPONENT_SCORE
-        for score in (
-            brightness_score_100,
-            contrast_score_100,
-            sharpness_score_100,
-            noise_score_100,
-        )
+    component_pass = (
+        brightness_score_100 > QUALITY_MIN_COMPONENT_SCORE
+        and contrast_score_100 > QUALITY_MIN_COMPONENT_SCORE
+        and sharpness_score_100 > QUALITY_MIN_SHARPNESS_SCORE
+        and noise_score_100 > QUALITY_MIN_COMPONENT_SCORE
     )
     advice = _quality_advice(
         brightness_mean=brightness_mean,
@@ -888,6 +888,9 @@ def compute_image_quality_metrics(img_bgr: np.ndarray) -> Dict[str, Any]:
         "quality_score_max": 100,
         "quality_accept_score": QUALITY_ACCEPT_SCORE,
         "quality_min_component_score": QUALITY_MIN_COMPONENT_SCORE,
+        "quality_min_sharpness_score": QUALITY_MIN_SHARPNESS_SCORE,
+        "quality_warning_component_score": QUALITY_WARNING_COMPONENT_SCORE,
+        "quality_warning_sharpness_score": QUALITY_WARNING_SHARPNESS_SCORE,
         "quality_pass": quality_score >= QUALITY_ACCEPT_SCORE and component_pass,
         "quality_advice": advice,
     }