import numpy as np '''Numpy and PIL implementation of a Mertens Fusion alghoritm Usage: Instantiate then set attributes: input_image = List containing path strings including .jpg Extension output_path = String ot Output without jpg ending compress_quality = 0-100 Jpeg compression level defaults to 75 Run function sequence() to start processing. Example: hdr = numpyHDR.NumpyHDR() hdr.input_image = photos/EV- stages/ hdr.compress_quality = 50 hdr.output_path = photos/result/ hdr.sequence() returns: Nothing ''' def simple_clip(fused,gamma): # Apply gamma correction #fused = np.clip(fused, 0, 1) fused = np.power(fused, 1.0 / gamma) #hdr_8bit = np.clip(res_mertens * 255, 0, 255).astype('uint8') fused = (255.0 * fused).astype(np.uint8) #fused = Image.fromarray(fused) return fused def convolve2d(image, kernel): # Get the dimensions of the input image and kernel image_height, image_width = image.shape kernel_height, kernel_width = kernel.shape # 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') # Define generators for row and column indices row_indices = range(image_height) col_indices = range(image_width) # Define a generator expression to generate patches centered at each pixel patches = ( padded_image[ row : row + kernel_height, col : col + kernel_width ] for row in row_indices for col in col_indices ) # Define a generator expression to generate element-wise products of patches and flipped kernels products = ( patch * np.flip(kernel, axis=(0, 1)) for patch in patches ) # Define a generator expression to generate convolved values convolved_values = ( product.sum() for product in products ) # Reshape the convolved values into an output image convolved_image = np.array(list(convolved_values)).reshape((image_height, image_width)) return convolved_image def mask(img, center=50, width=20, threshold=0.2): '''Mask with sigmoid smooth''' mask = 1 / (1 + np.exp((center - img) / width)) # Smooth gradient mask mask = np.where(img > threshold, mask, 1) # Apply threshold to the mask mask = img * mask #plot_histogram(mask, title="mask") return mask def highlightsdrop(img, center=0.7, width=0.2, threshold=0.6, amount=0.08): '''Mask with sigmoid smooth targets bright sections''' mask = 1 / (1 + np.exp((center - img) / width)) # Smooth gradient mask mask = np.where(img > threshold, mask, 0) # Apply threshold to the mask mask = mask.reshape((img.shape)) print(np.max(mask)) img_adjusted = img - (mask * amount) # Adjust the image with a user-specified amount img_adjusted = np.clip(img_adjusted, 0, 1) return img_adjusted def shadowlift(img, center=0.2, width=0.1, threshold=0.2, amount= 0.05): '''Mask with sigmoid smooth targets bright sections''' mask = 1 / (1 + np.exp((center - img) / width)) # Smooth gradient mask mask = np.where(img < threshold, mask, 0) # Apply threshold to the mask mask = mask.reshape((img.shape)) print(np.max(mask)) img_adjusted = (mask * amount) + img # Adjust the image with a user-specified amount img_adjusted = np.clip(img_adjusted, 0, 1) return img_adjusted def blur(image, amount=1): # Define a kernel for sharpening kernel = np.array([[0, -1, 0], [-1, 4, -1], [0, -1, 0]]) # Apply the kernel to each channel of the image using convolution blurred = convolve2d(image, kernel) # Add the original image to the sharpened image with a weight of the sharpening amount sharpened = image + amount * (image - blurred) sharpened = np.clip(sharpened, 0, 1) # Crop the output image to match the input size #sharpened = sharpened.reshape(image.shape) return sharpened def mertens_fusion(stack, gamma:float =1, contrast_weight:float =1 ,blurred: bool = False) -> bytearray: """Fuse multiple exposures into a single HDR image using the Mertens algorithm. Args: image_paths: A list of paths to input images. gamma: The gamma correction value to apply to the input images. contrast_weight: The weight of the local contrast term in the weight map computation. blurred: Helps making transitions for the weights smoother but increases provessing time x2 Returns: The fused HDR image. """ images = [] for array in stack: #Incoming arrays in 255 er range img = np.array(array).astype(np.float32) / 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 = [] for img in images: threshold_h = .99 threshold_l = .1 # Apply thresholding to filter out overexposed portions of the image img = np.where(img > threshold_h, 0.99, img) gray = np.dot(img, [0.2989, 0.5870, 0.1140]) 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)) weight = np.power(laplacian, contrast_weight) weight_maps.append(weight) # Normalize the weight maps. total_weight = sum(weight_maps) 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) for i, img in enumerate(images): fused += weight_maps[i][:, :, np.newaxis] * img return fused def compress_dynamic_range(image): '''Compress dynamic range based on percentile''' # Find the 1st and 99th percentiles of the image p1, p99 = np.percentile(image, (0, 99)) # Calculate the range of the image img_range = p99 - p1 # Calculate the compression factor required to fit the image into 8-bit range c = 1 / img_range # Subtract the 1st percentile from the image and clip it to the [0, 1] range new_image = np.clip((image - p1) * c, 0, 1) return new_image def compress_dynamic_range_histo(image, new_min=0.01, new_max=0.99): """Compress the dynamic range of an image using histogram stretching. Args: image: A numpy array representing an image. new_min: The minimum value of the new range. new_max: The maximum value of the new range. Returns: The compressed image. """ # Calculate the histogram of the image. hist, bins = np.histogram(image.ravel(), bins=256, range=(0, 1)) # Calculate the cumulative distribution function (CDF) of the histogram. cdf = hist.cumsum() cdf = (cdf - cdf.min()) / (cdf.max() - cdf.min()) # normalize to [0, 1] # Interpolate the CDF to get the new pixel values. new_pixels = np.interp(image.ravel(), bins[:-1], cdf * (new_max - new_min) + new_min) # Reshape the new pixel values to the shape of the original image. new_image = new_pixels.reshape((image.shape[0], image.shape[1], image.shape[2])) return new_image def process(stack, gain: float = 1, weight: float = 1, gamma: float = 1, post: bool = True, blurred: bool = True): '''Processes the stack that contains a list of arrays form the camera into a PIL compatible clipped output array Args: stack : input list with arrays gain : low value low contrast, high value high contrast and brightness weight: How much the extracted portions of each image gets allpied to to the result "HDR effect intensity" gamma: Post fusion adjustment of the gamma. post: Enable or disable effects applied after the fusion True or False, default True shadowlift = slightly lifts the shadows Args: center: position of the filter dropoff width: range of the gradient, softness threshold: sets the threshhold form 0 to 1 0.1= lowest blacks.... amount: How much the shadows should be lifted. Values under 0.1 seem to be good. returns: Hdr image with lifted blacks clipped to 0,1 range compress dynamic range: Tries to fit the image better into the available range. Less loggy image. Returns: HDR Image as PIL compatible array. ''' hdr_image = mertens_fusion(stack ,gain, weight, blurred) if post == True: #hdr_image = self.highlightsdrop(hdr_image) hdr_image = shadowlift(hdr_image) hdr_image = compress_dynamic_range(hdr_image) #hdr_image = self.compress_dynamic_range_histo(hdr_image) hdr_image = simple_clip(hdr_image,gamma) return hdr_image