MohsinElis commited on
Commit
11d1bb5
·
verified ·
1 Parent(s): 9d174ec

perf: use pure vectorized numpy for blockiness and saturation

Browse files
Files changed (1) hide show
  1. api/model_server.py +23 -19
api/model_server.py CHANGED
@@ -88,14 +88,14 @@ def check_photographic_domain(
88
  / total_pixels
89
  )
90
 
91
- # Color saturation in HSV
92
- try:
93
- import cv2
94
- hsv = cv2.cvtColor(np.array(img_rgb), cv2.COLOR_RGB2HSV)
95
- sat = hsv[:, :, 1]
96
- high_sat_ratio = float(np.mean(sat > 100))
97
- except Exception:
98
- high_sat_ratio = 0.0
99
 
100
  is_ood = False
101
  domain_status = "photographic"
@@ -146,18 +146,22 @@ def check_photographic_domain(
146
  f"Non-photographic artwork."
147
  )
148
 
149
- # 8x8 DCT JPEG Blockiness calculation
 
 
150
  blockiness = 1.0
151
- try:
152
- import cv2
153
- gray = cv2.cvtColor(np.array(img_rgb), cv2.COLOR_RGB2GRAY).astype(np.float32)
154
- h_blocks = h // 8
155
- if h_blocks > 2 and w > 16:
156
- boundary_diff = np.mean([np.mean(np.abs(gray[i * 8, :] - gray[i * 8 - 1, :])) for i in range(1, h_blocks)])
157
- interior_diff = np.mean([np.mean(np.abs(gray[i * 8 - 1, :] - gray[i * 8 - 2, :])) for i in range(1, h_blocks)])
158
- blockiness = float(boundary_diff / (interior_diff + 1e-6))
159
- except Exception:
160
- blockiness = 1.0
 
 
161
 
162
  # Social media compression indicators
163
  social_media_keywords = [
 
88
  / total_pixels
89
  )
90
 
91
+ # Pure NumPy Grayscale (Rec. 601 luma)
92
+ gray = 0.299 * arr[:, :, 0] + 0.587 * arr[:, :, 1] + 0.114 * arr[:, :, 2]
93
+
94
+ # Color saturation in pure NumPy
95
+ max_c = np.max(arr, axis=2)
96
+ min_c = np.min(arr, axis=2)
97
+ sat = np.where(max_c > 0, (max_c - min_c) / (max_c + 1e-6) * 255.0, 0.0)
98
+ high_sat_ratio = float(np.mean(sat > 100))
99
 
100
  is_ood = False
101
  domain_status = "photographic"
 
146
  f"Non-photographic artwork."
147
  )
148
 
149
+ # 8x8 DCT JPEG Blockiness calculation (Pure NumPy, vectorized across horizontal and vertical axes)
150
+ h_blocks = h // 8
151
+ w_blocks = w // 8
152
  blockiness = 1.0
153
+ if h_blocks > 2 and w_blocks > 2:
154
+ b_rows = np.arange(1, h_blocks) * 8
155
+ boundary_diff_h = np.mean(np.abs(gray[b_rows, :] - gray[b_rows - 1, :]))
156
+ interior_diff_h = np.mean(np.abs(gray[b_rows - 1, :] - gray[b_rows - 2, :]))
157
+ blk_h = boundary_diff_h / (interior_diff_h + 1e-6)
158
+
159
+ b_cols = np.arange(1, w_blocks) * 8
160
+ boundary_diff_v = np.mean(np.abs(gray[:, b_cols] - gray[:, b_cols - 1]))
161
+ interior_diff_v = np.mean(np.abs(gray[:, b_cols - 1] - gray[:, b_cols - 2]))
162
+ blk_v = boundary_diff_v / (interior_diff_v + 1e-6)
163
+
164
+ blockiness = float((blk_h + blk_v) / 2.0)
165
 
166
  # Social media compression indicators
167
  social_media_keywords = [