| |
| """Perceptual video hash: one 64-bit dHash per second, compared with time-shift search. |
| |
| usage: python3 vhash.py a.mp4 b.mp4 [c.mp4 ...] |
| Prints a distance matrix (average differing bits out of 64, lower = more alike). |
| """ |
| import subprocess, sys |
| import numpy as np |
|
|
| def dhashes(path, step=1.0): |
| """dHash of one frame per `step` seconds (9x8 gray, compare neighbours).""" |
| cmd = ["ffmpeg", "-v", "error", "-i", path, "-vf", |
| f"fps=1/{step}:start_time=0.5,scale=9:8:flags=area,format=gray", |
| "-f", "rawvideo", "-"] |
| raw = subprocess.run(cmd, capture_output=True, check=True).stdout |
| px = np.frombuffer(raw, np.uint8).reshape(-1, 8, 9).astype(np.int16) |
| return (px[:, :, 1:] > px[:, :, :-1]).reshape(len(px), 64) |
|
|
| def distance(a, b, max_shift=3): |
| """Best mean Hamming distance over whole-second shifts of b against a.""" |
| best = (64.0, 0) |
| for s in range(-max_shift, max_shift + 1): |
| x = a[max(0, s):] |
| y = b[max(0, -s):] |
| n = min(len(x), len(y)) |
| if n < 4: |
| continue |
| d = float(np.mean(np.sum(x[:n] != y[:n], axis=1))) |
| best = min(best, (d, s)) |
| return best |
|
|
| if __name__ == "__main__": |
| files = sys.argv[1:] |
| H = [dhashes(f) for f in files] |
| names = [f.rsplit("/", 1)[-1][:14] for f in files] |
| print(" " * 16 + "".join(f"{n[:9]:>11}" for n in names)) |
| for i, a in enumerate(H): |
| cells = [] |
| for j, b in enumerate(H): |
| d, s = distance(a, b) |
| cells.append(f"{d:5.1f}({s:+d})") |
| print(f"{names[i]:<16}" + "".join(f"{c:>11}" for c in cells)) |
|
|
