傳統雜湊 (MD5/SHA) 對輸入極敏感,單位元組變化導致完全不同輸出。**感知雜湊** 根據圖像視覺內容生成指紋,相似圖片產生相似雜湊 (漢明距離小),適用於去重、版權偵測、內容審核。
pip install pillow imagehash numpy
import argparse
import shelve
from PIL import Image
import imagehash
from pathlib import Path
def build_index(dataset_path, db_path):
db = shelve.open(db_path, writeback=True)
image_paths = list(Path(dataset_path).rglob('*.[jp][pn]g'))
print(f"處理 {len(image_paths)} 張圖片...")
for i, path in enumerate(image_paths):
try:
with Image.open(path) as img:
# 計算 dHash
hash_val = imagehash.dhash(img)
# 儲存:hash(str) -> 路徑列表
key = str(hash_val)
if key not in db:
db[key] = []
db[key].append(str(path))
except Exception as e:
print(f"跳過 {path}: {e}")
if (i+1) % 100 == 0:
print(f" 已處理 {i+1}/{len(image_paths)}")
db.sync()
db.close()
print(f"索引建立完成,共 {len(db)} 個唯一指紋")
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--dataset', required=True, help='圖片資料集路徑')
parser.add_argument('-s', '--shelve', required=True, help='輸出資料庫路徑')
args = parser.parse_args()
build_index(args.dataset, args.shelve)
def search_similar(query_image_path, db_path, threshold=10):
"""搜尋漢明距離 <= threshold 的圖片"""
db = shelve.open(db_path, flag='r')
with Image.open(query_image_path) as img:
query_hash = imagehash.dhash(img)
results = []
for hash_str, paths in db.items():
db_hash = imagehash.hex_to_hash(hash_str)
distance = query_hash - db_hash # imagehash 重載運算子
if distance <= threshold:
for p in paths:
results.append((distance, p))
db.close()
return sorted(results)
# 使用
results = search_similar('test.jpg', 'index.db', threshold=15)
for dist, path in results[:10]:
print(f"距離 {dist:2d}: {path}")
線性掃描資料庫隨資料增長極慢。度量空間索引結構:
# 使用 pybktree (pip install pybktree)
from pybktree import BKTree
def build_bktree(db_path):
db = shelve.open(db_path, flag='r')
tree = BKTree(imagehash.hex_to_hash, list(db.keys()))
db.close()
return tree
tree = build_bktree('index.db')
query_hash = imagehash.dhash(Image.open('query.jpg'))
# 查詢距離 <= 10
matches = tree.find(query_hash, 10)
for dist, hash_str in matches:
print(f"距離 {dist}: {db[hash_str]}")
文中所述真實案例:建立違規圖片指紋庫,用戶上傳頭像時即時比對,阻擋 80%+ 違規上傳。關鍵點:
| 算法 | 特點 | 適用場景 |
|---|---|---|
| aHash (平均雜湊) | 最簡單,縮放 8x8 比較均值 | 縮圖識別、極簡場景 |
| pHash (感知雜湊) | DCT 頻域,抗縮放/壓縮/旋轉 | 版權保護、視頻幀檢索 |
| dHash (差值雜湊) | 梯度資訊,速度快、精度好 | 通用相似搜尋、去重 |
| wHash (小波雜湊) | 小波變換,多尺度特徵 | 複雜紋理、醫學影像 |
imagehash 庫支援全部:imagehash.average_hash、phash、dhash、whash。