目标检测讲解

news2025/2/27 17:32:29

环境准备

pip install scikit-image -i https://pypi.tuna.tsinghua.edu.cn/simple

图片读取&画框

from skimage import io
import matplotlib.pyplot as plt
import matplotlib.patches as mp

ss = io.imread('dogs.jpg')

_, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(ss)

rect = mp.Rectangle((600, 400), 100, 80, fill=False, edgecolor='red', linewidth=1)
ax.add_patch(rect)

plt.show()

运行代码，显示如下

Selective Search

方法特点

捕捉不同尺度（Capture All Scales)
多样化（Diversification)
快速计算（Fast to Compute）

图像中区域特征比像素更具代表性

计算所有邻近区域之间的相似性；
两个最相似的区域被组合在一起；
计算合并区域和相邻区域的相似度；
重复2、3过程，直到整个图像变为一个区域

另外加入一些多样性的衡量

颜色相似性
纹理相似性
尺度相似性
形状重合度

效果如下

Selective Search环境准备

pip install selectivesearch

生成示例图片的框

from skimage import io
import selectivesearch

ss = io.imread('dogs.jpg')

# ss生成目标框
_, regions = selectivesearch.selective_search(ss, scale=500, sigma=0.9, min_size=10)

print(len(regions))
for i in regions[0:5]:
    print(i)

运行代码，打印信息如下

2437
{'rect': (0, 0, 1299, 951), 'size': 47450, 'labels': [0.0]}
{'rect': (48, 0, 96, 0), 'size': 97, 'labels': [1.0]}
{'rect': (432, 0, 319, 0), 'size': 320, 'labels': [2.0]}
{'rect': (1120, 0, 47, 1), 'size': 64, 'labels': [3.0]}
{'rect': (0, 7, 340, 0), 'size': 341, 'labels': [4.0]}

对候选框进行过滤

去除重复框
删除小框
删除宽高比较大的框

代码如下

candidates = set()
for r in regions:
    # with different segments
    if r['rect'] in candidates:
        continue
    # 删除 < 2000 pixels
    if r['size'] < 2000:
        continue
    # 删除 宽高比 > 1.2
    x, y, w, h = r['rect']
    if w / h > 1.2 or h / w > 1.2:
        continue
    candidates.add(r['rect'])
print(candidates)

运行代码，打印如下

{(857, 523, 429, 375), (864, 520, 426, 431), (207, 266, 139, 143), (508, 523, 330, 373), (1058, 344, 75, 65), (75, 522, 405, 375), (854, 520, 434, 378), (1060, 344, 73, 64), (23, 32, 233, 268), (75, 522, 406, 375), (176, 600, 307, 298), (1059, 11, 141, 165), (0, 133, 258, 275), (493, 523, 346, 376), (1105, 285, 141, 124), (1102, 280, 144, 129), (855, 521, 431, 377), (776, 8, 519, 503), (396, 122, 279, 285), (1059, 344, 74, 64), (1101, 236, 137, 129), (508, 523, 330, 376), (137, 551, 347, 347), (508, 523, 331, 376), (65, 521, 419, 377), (589, 228, 172, 181), (1057, 11, 93, 88), (0, 8, 349, 401), (1101, 227, 186, 182), (154, 551, 330, 347), (855, 521, 433, 377), (688, 750, 150, 148), (1129, 285, 117, 124), (1057, 9, 229, 216), (1059, 11, 142, 165), (363, 9, 399, 400), (23, 32, 323, 377), (363, 9, 398, 400), (341, 8, 423, 405), (176, 551, 308, 347), (887, 650, 186, 187), (176, 600, 308, 298), (23, 32, 324, 377)}

显示满足条件的候选框

from skimage import io
import selectivesearch
import matplotlib.pyplot as plt
import matplotlib.patches as mp

ss = io.imread('dogs.jpg')

_, regions = selectivesearch.selective_search(ss, scale=500, sigma=0.9, min_size=10)

# 过滤框
candidates = set()
index = 0
for r in regions:
    index += 1
    print(index)
    # with different segments
    if r['rect'] in candidates:
        continue
    # 删除 < 2000 pixels
    if r['size'] < 2000:
        continue
    # 删除 宽高比 > 1.2
    x, y, w, h = r['rect']
    if w / h > 1.2 or h / w > 1.2:
        continue
    candidates.add(r['rect'])

# 画图
fig, ax = plt.subplots(ncols=1, nrows=1, figsize=(6, 6))
ax.imshow(ss)
for x, y, w, h in candidates:
    rect = mp.Rectangle((x, y), w, h, fill=False, edgecolor='red', linewidth=1)
    ax.add_patch(rect)

plt.show()