"""Homework 02: Filtering. Complete the three TODO functions below. Do not change their names or arguments; the Gradescope autograder calls them directly. Run locally from the course `cv` Conda environment with: python hw02.py my_image.png The script will save a visualization to hw02_output.png. You ONLY submit hw02.py to Gradescope. """ from __future__ import annotations import argparse from pathlib import Path import cv2 import numpy as np BOX_3X3 = np.ones((3, 3), dtype=np.float32) / 9.0 SOBEL_X = np.array( [ [-1, 0, 1], [-2, 0, 2], [-1, 0, 1], ], dtype=np.float32, ) SHARPEN_3X3 = np.array( [ [0, -1, 0], [-1, 5, -1], [0, -1, 0], ], dtype=np.float32, ) def correlate2d(image: np.ndarray, kernel: np.ndarray) -> np.ndarray: """Return same-size 2-D correlation with zero padding. Requirements: - `image` and `kernel` are 2-D arrays. - The kernel has odd height and width. - Use zero padding so the output has the same shape as the input. - Apply the kernel as written; do NOT rotate it. - Implement the neighborhood sweep yourself using nested `for` loops. - Do not call a library filtering/correlation/convolution routine. - Return a NumPy array with dtype np.float32. """ # TODO raise NotImplementedError def convolve2d(image: np.ndarray, kernel: np.ndarray) -> np.ndarray: """Return same-size 2-D convolution with zero padding. Rotate the kernel by 180 degrees, then reuse correlate2d(). You may use np.flip() or cv2.flip() to rotate the kernel; the filtering itself must still be performed by your correlate2d() implementation. """ # TODO raise NotImplementedError def median_filter3x3(image: np.ndarray) -> np.ndarray: """Return a same-size 3x3 median-filtered image with zero padding. Sweep the image yourself using nested `for` loops. You may use np.median() to compute the median of each local 3x3 neighborhood, but you may NOT use cv2.medianBlur(), scipy.ndimage.median_filter(), or another library median filtering implementation. Return dtype np.float32. """ # TODO raise NotImplementedError # ----------------------------------------------------------------------------- # The helper functions below are provided. You do not need to modify them. # ----------------------------------------------------------------------------- def load_grayscale_float(filename: str | Path) -> np.ndarray: """Load an image with OpenCV and return grayscale float32 in [0, 1].""" image = cv2.imread(str(filename), cv2.IMREAD_GRAYSCALE) if image is None: raise FileNotFoundError(f"Could not load image: {filename}") return image.astype(np.float32) / 255.0 def resize_for_filtering(image: np.ndarray, max_dimension: int = 96) -> np.ndarray: """Downsize large images so the manual Python loops run quickly.""" height, width = image.shape largest = max(height, width) if largest <= max_dimension: return image scale = max_dimension / largest new_size = (max(1, round(width * scale)), max(1, round(height * scale))) return cv2.resize(image, new_size, interpolation=cv2.INTER_AREA) def add_salt_and_pepper_noise( image: np.ndarray, probability: float = 0.08, seed: int = 7 ) -> np.ndarray: """Create deterministic salt-and-pepper noise for the visualization.""" rng = np.random.default_rng(seed) noisy = image.copy() random_values = rng.random(image.shape) noisy[random_values < probability / 2] = 0.0 noisy[random_values > 1.0 - probability / 2] = 1.0 return noisy.astype(np.float32) def _to_display_u8(image: np.ndarray, signed_response: bool = False) -> np.ndarray: """Convert a float image into a viewable uint8 image for the montage.""" arr = np.asarray(image, dtype=np.float32) if signed_response: arr = np.abs(arr) lo = float(np.min(arr)) hi = float(np.max(arr)) if hi - lo < 1e-12: scaled = np.zeros_like(arr, dtype=np.float32) elif lo >= 0.0 and hi <= 1.0: scaled = arr else: scaled = (arr - lo) / (hi - lo) return np.clip(np.rint(scaled * 255.0), 0, 255).astype(np.uint8) def save_montage( panels: list[tuple[str, np.ndarray, bool]], output_path: str | Path = "hw02_output.png", ) -> None: """Save a 2x3 OpenCV montage of filtering results.""" rendered = [] for label, image, signed_response in panels: panel = _to_display_u8(image, signed_response) panel = cv2.cvtColor(panel, cv2.COLOR_GRAY2BGR) cv2.rectangle(panel, (0, 0), (panel.shape[1] - 1, 20), (255, 255, 255), -1) cv2.putText( panel, label, (4, 15), cv2.FONT_HERSHEY_SIMPLEX, 0.42, (0, 0, 0), 1, cv2.LINE_AA, ) rendered.append(panel) row1 = cv2.hconcat(rendered[:3]) row2 = cv2.hconcat(rendered[3:6]) montage = cv2.vconcat([row1, row2]) if not cv2.imwrite(str(output_path), montage): raise RuntimeError(f"Could not save {output_path}") def main() -> None: parser = argparse.ArgumentParser(description="HW02: Filtering") parser.add_argument("image", help="Path to a JPG/PNG image (you may reuse your HW01 headshot)") args = parser.parse_args() image = resize_for_filtering(load_grayscale_float(args.image)) box_corr = correlate2d(image, BOX_3X3) box_conv = convolve2d(image, BOX_3X3) sobel_corr = correlate2d(image, SOBEL_X) sobel_conv = convolve2d(image, SOBEL_X) sharpened = correlate2d(image, SHARPEN_3X3) noisy = add_salt_and_pepper_noise(image) median = median_filter3x3(noisy) print( "Box correlation vs convolution max difference:", float(np.max(np.abs(box_corr - box_conv))), ) print( "Sobel-X correlation vs convolution max difference:", float(np.max(np.abs(sobel_corr - sobel_conv))), ) save_montage( [ ("Original", image, False), ("Box 3x3", box_corr, False), ("|Sobel X|", sobel_corr, True), ("Sharpen", sharpened, False), ("Salt & pepper", noisy, False), ("Median 3x3", median, False), ], "hw02_output.png", ) print("Saved visualization to hw02_output.png") if __name__ == "__main__": main()