Reduce build size by dropping scipy in favor of opencv-contrib-python-headless

pyproject.toml

@@ -9,18 +9,12 @@ dependencies = [
   "PyWinCtl >=0.0.42", # py.typed
   "keyboard @ git+https://github.com/boppreh/keyboard.git", # Fix install on macos and linux-ci https://github.com/boppreh/keyboard/pull/568
   "numpy >=2.3", # Windows ARM64 wheels
+  "opencv-contrib-python-headless >=4.10", # NumPy 2 support
   "packaging >=20.0", # py.typed
   # When needed, dev builds can be found at https://download.qt.io/snapshots/ci/pyside/dev?C=M;O=D
   "PySide6-Essentials", # Let package resolution find the minimum for wheels (>=6.9.0 on Windows ARM64; <6.8.1 on ubuntu-22.04-arm (glibc 2.39))
   "tomli-w >=1.1.0", # Typing fixes
 
-  # scipy is used for pHash calculation as a smaller, but still fast implementation.
-  # However, scipy is not available on all environments yet.
-  # In those cases, we're falling back to opencv-contrib-python's cv2.img_hash
-  "opencv-contrib-python-headless; platform_machine == 'ARM64' or platform_machine == 'aarch64'",
-  "opencv-python-headless; platform_machine != 'ARM64' and platform_machine != 'aarch64'",
-  "scipy >=1.14.1; platform_machine != 'ARM64' and platform_machine != 'aarch64'", # Python 3.13 support
-
   #
   # Build and compile resources
   "pyinstaller >=6.14.0", # Mitigate issues with pkg_resources deprecation warning
@@ -59,7 +53,6 @@ dev = [
   "ruff >=0.11.13",
   #
   # Types
-  "scipy-stubs >=1.14.1.1",
   "types-PyAutoGUI",
   "types-PyScreeze; sys_platform == 'linux'",
   "types-keyboard",

src/compare.py

@@ -1,6 +1,5 @@
 from collections.abc import Iterable
 from math import sqrt
-from typing import TYPE_CHECKING
 
 import cv2
 import Levenshtein
@@ -22,6 +21,7 @@
 RANGES = (0, MAXRANGE, 0, MAXRANGE, 0, MAXRANGE)
 MASK_SIZE_MULTIPLIER = ColorChannel.Alpha * MAXBYTE * MAXBYTE
 MAX_VALUE = 1.0
+CV2_PHASH_SIZE = 8
 
 
 def compare_histograms(source: MatLike, capture: MatLike, mask: MatLike | None = None):
@@ -90,41 +90,39 @@ def compare_template(source: MatLike, capture: MatLike, mask: MatLike | None = N
     return 1 - (min_val / max_error)
 
 
-try:
-    from scipy import fft
-
-    def __cv2_scipy_compute_phash(image: MatLike, hash_size: int, highfreq_factor: int = 4):
-        """Implementation copied from https://github.com/JohannesBuchner/imagehash/blob/38005924fe9be17cfed145bbc6d83b09ef8be025/imagehash/__init__.py#L260 ."""  # noqa: E501
-        img_size = hash_size * highfreq_factor
-        image = cv2.cvtColor(image, cv2.COLOR_BGRA2GRAY)
-        image = cv2.resize(image, (img_size, img_size), interpolation=cv2.INTER_AREA)
-        dct = fft.dct(fft.dct(image, axis=0), axis=1)
-        dct_low_frequency = dct[:hash_size, :hash_size]
-        median = np.median(dct_low_frequency)
-        return dct_low_frequency > median
-
-    def __cv2_phash(source: MatLike, capture: MatLike, hash_size: int = 8):
-        source_hash = __cv2_scipy_compute_phash(source, hash_size)
-        capture_hash = __cv2_scipy_compute_phash(capture, hash_size)
-        hash_diff = np.count_nonzero(source_hash != capture_hash)
-        return 1 - (hash_diff / 64.0)
-
-except ModuleNotFoundError:
-    if not TYPE_CHECKING:  # opencv-contrib-python-headless being installed is based on architecture
-
-        def __cv2_phash(source: MatLike, capture: MatLike, hash_size: int = 8):
-            # OpenCV has its own pHash comparison implementation in `cv2.img_hash`,
-            # but it requires contrib/extra modules and is inaccurate
-            # unless we precompute the size with a specific interpolation.
-            # See: https://github.com/opencv/opencv_contrib/issues/3295#issuecomment-1172878684
-            #
-            phash = cv2.img_hash.PHash.create()
-            source = cv2.resize(source, (hash_size, hash_size), interpolation=cv2.INTER_AREA)
-            capture = cv2.resize(capture, (hash_size, hash_size), interpolation=cv2.INTER_AREA)
-            source_hash = phash.compute(source)
-            capture_hash = phash.compute(capture)
-            hash_diff = phash.compare(source_hash, capture_hash)
-            return 1 - (hash_diff / 64.0)
+# The old scipy-based implementation.
+# Turns out this cuases an extra 25 MB build compared to opencv-contrib-python-headless
+# # from scipy import fft
+# def __cv2_scipy_compute_phash(image: MatLike, hash_size: int, highfreq_factor: int = 4):
+#     """Implementation copied from https://github.com/JohannesBuchner/imagehash/blob/38005924fe9be17cfed145bbc6d83b09ef8be025/imagehash/__init__.py#L260 ."""  # noqa: E501
+#     img_size = hash_size * highfreq_factor
+#     image = cv2.cvtColor(image, cv2.COLOR_BGRA2GRAY)
+#     image = cv2.resize(image, (img_size, img_size), interpolation=cv2.INTER_AREA)
+#     dct = fft.dct(fft.dct(image, axis=0), axis=1)
+#     dct_low_frequency = dct[:hash_size, :hash_size]
+#     median = np.median(dct_low_frequency)
+#     return dct_low_frequency > median
+# def __cv2_phash(source: MatLike, capture: MatLike, hash_size: int = 8):
+#     source_hash = __cv2_scipy_compute_phash(source, hash_size)
+#     capture_hash = __cv2_scipy_compute_phash(capture, hash_size)
+#     hash_diff = np.count_nonzero(source_hash != capture_hash)
+#     return 1 - (hash_diff / 64.0)
+
+
+def __cv2_phash(source: MatLike, capture: MatLike):
+    """
+    OpenCV has its own pHash comparison implementation in `cv2.img_hash`,
+    but is inaccurate unless we precompute the size with a specific interpolation.
+
+    See: https://github.com/opencv/opencv_contrib/issues/3295#issuecomment-1172878684
+    """
+    phash = cv2.img_hash.PHash.create()
+    source = cv2.resize(source, (CV2_PHASH_SIZE, CV2_PHASH_SIZE), interpolation=cv2.INTER_AREA)
+    capture = cv2.resize(capture, (CV2_PHASH_SIZE, CV2_PHASH_SIZE), interpolation=cv2.INTER_AREA)
+    source_hash = phash.compute(source)
+    capture_hash = phash.compute(capture)
+    hash_diff = phash.compare(source_hash, capture_hash)
+    return 1 - (hash_diff / 64.0)
 
 
 def compare_phash(source: MatLike, capture: MatLike, mask: MatLike | None = None):