Integrated option for compiling to cython

convolve2d_cython.pyx

@@ -0,0 +1,42 @@
+# Import necessary packages
+import cython
+import numpy as np
+cimport numpy as np
+
+
+# Declare types for function arguments and variables
+cpdef np.ndarray[np.float64_t, ndim=2] convolve2d(np.ndarray[np.float64_t, ndim=2] image,
+                                                  np.ndarray[np.float64_t, ndim=2] kernel):
+
+    cdef int image_height, image_width, kernel_height, kernel_width, pad_height, pad_width, row, col
+    cdef np.ndarray[np.float64_t, ndim=2] padded_image, convolved_image, patch, product
+
+    # Get the dimensions of the input image and kernel
+    # Get the dimensions of the input image and kernel
+    image_height, image_width = int(image.shape[0]), int(image.shape[1])
+    kernel_height, kernel_width = int(kernel.shape[0]), int(kernel.shape[1])
+
+    # Compute the padding needed to handle boundary effects
+    pad_height = (kernel_height - 1) // 2
+    pad_width = (kernel_width - 1) // 2
+    padded_image = np.pad(image, ((pad_height, pad_height), (pad_width, pad_width)), mode='constant')
+
+    # Initialize the output image
+    convolved_image = np.zeros((image_height, image_width), dtype=np.float64)
+
+    # Loop over each pixel in the output image and compute the convolved value
+    for row in range(image_height):
+        for col in range(image_width):
+            # Extract the patch centered at the current pixel
+            patch = padded_image[row : row + kernel_height, col : col + kernel_width]
+
+            # Compute the element-wise product of the patch and the flipped kernel
+            product = patch * np.flip(kernel, axis=(0, 1))
+
+            # Compute the sum of the element-wise products
+            convolved_value = np.sum(product)
+
+            # Store the convolved value in the output image
+            convolved_image[row, col] = convolved_value
+
+    return convolved_image

makehdr.py

@@ -1,11 +1,13 @@
 #!//usr/bin/python3
 import file_utility as file
 import numpyHDR as hdr
+import os
 
 '''CLI application for HDR experiments'''
 
+
 stack = []
-select = input("Select Image Source: 1 - Raspicam, 2 - From File, 3 - Image sequence, 4 debug: ")
+select = input("Select Image Source: 1 - Raspicam, 2 - From File, 3 - Image sequence, 4 debug, 5 - compile Cython: ")
 
 if int(select) == 1:
     import picamburst as pcb
@@ -36,7 +38,15 @@ if int(select) == 4:
     path_list = ['webcam25_3_2023_ev0.jpg','webcam25_3_2023_ev1.jpg','webcam25_3_2023_ev2.jpg']
     stack = file.openImageList(path_list, True)
 
-print(path_list)
+if int(select) == 5:
+    try:
+        os.system('python3 setup.py build_ext --inplace')
+    except Exception as e:
+        print("Error while compiling cython function", e)
+    print("Please restart")
+    exit()
+
+#print(path_list)
 
 #Process HDR with mertens fusion and post effects, blur
 #Set last value to false for double the speed but lesser blancaed hdr effect.

numpyHDR.py

@@ -1,5 +1,15 @@
 import numpy as np
 
+try:
+    import convolve2d_cython
+    available = True
+    print("Using compiled Cython Convolve")
+except ImportError:
+    available = False
+    print("Using normal Numpy Convolve")
+
+
+
 '''Numpy and PIL implementation of a Mertens Fusion alghoritm
 Usage: Instantiate then set attributes:
     input_image = List containing path strings including .jpg Extension
@@ -103,8 +113,13 @@ def blur(image, amount=1):
                        [0, -1, 0]])
 
     # Apply the kernel to each channel of the image using convolution
-    blurred = convolve2d(image, kernel)
+    #blurred = convolve2d(image, kernel)
-
+    kernel = kernel.astype(np.float64)
+    #image= image.astype(np.float64)
+    if available:
+        blurred = convolve2d_cython.convolve2d(image, kernel)
+    else:
+        blurred = convolve2d(image, kernel)
     # Add the original image to the sharpened image with a weight of the sharpening amount
     sharpened = image + amount * (image - blurred)
 
@@ -133,12 +148,13 @@ def mertens_fusion(stack, gamma:float =1, contrast_weight:float =1 ,blurred: boo
     images = []
     for array in stack:
         #Incoming arrays in 255 er range
-        img = np.array(array).astype(np.float32) / 255.0
+        img = np.array(array).astype(np.float64) / 255.0
         img = np.power(img, gamma)
         images.append(img)
 
     # Compute the weight maps for each input image based on the local contrast.
     weight_maps = []
+    kernel = np.array([[1, 2, 1], [2, -11, 2], [1, 2, 1]])
 
     for img in images:
         threshold_h = .99
@@ -149,8 +165,12 @@ def mertens_fusion(stack, gamma:float =1, contrast_weight:float =1 ,blurred: boo
         if blurred:
             gray = blur(gray, 1)
         #kernel = np.array([[-1, -1, -1], [-1, 8, -1], [-1, -1, -1]])
-        kernel = np.array([[1, 2, 1], [2, -11, 2], [1, 2, 1]])
+
-        laplacian = np.abs(convolve2d(gray, kernel))
+        kernel = kernel.astype(np.float64)
+        if available:
+            laplacian = np.abs(convolve2d_cython.convolve2d(gray, kernel))
+        else:
+            laplacian = np.abs(convolve2d(gray, kernel))
         weight = np.power(laplacian, contrast_weight)
         weight_maps.append(weight)
 
@@ -159,7 +179,7 @@ def mertens_fusion(stack, gamma:float =1, contrast_weight:float =1 ,blurred: boo
     weight_maps = [w / total_weight for w in weight_maps]
 
     # Compute the fused HDR image by computing a weighted sum of the input images.
-    fused = np.zeros(images[0].shape, dtype=np.float32)
+    fused = np.zeros(images[0].shape, dtype=np.float64)
     for i, img in enumerate(images):
         fused += weight_maps[i][:, :, np.newaxis] * img
 

setup.py

@@ -0,0 +1,9 @@
+from setuptools import setup
+from Cython.Build import cythonize
+import numpy
+
+setup(
+    name='convolve2d_cython',
+    ext_modules=cythonize("convolve2d_cython.pyx"),
+    include_dirs=[numpy.get_include()]
+)