机器学习: 性能度量

# coding=utf-8

import sys
reload(sys)
sys.setdefaultencoding('utf-8')

... ...

def performance(tp, fn, fp, tn):
    # 查准率
    P = tp / float(tp + fp)
    # 查全率
    R = tp / float(tp + fn)
    # F1
    F1 = (2 * P * R) / (P + R)
    # print P, R, F1
    return P, R, F1

... ...

P, R, F1 = performance(white_pixels, red_pixels, green_pixels, black_pixels)
print '查准率 P = {:>.3f},  查全率 R = {:>.3f},   F1 = {:>.3f}'.format(P, R, F1).encode('gb18030')

... ...

# coding=utf-8

import sys
reload(sys)
sys.setdefaultencoding('utf-8')

diff_path = './data/diff.jpg'

black_pixels, white_pixels, green_pixels, red_pixels = 29158, 530899, 75994, 3949
sum_pixels = 640000

def performance(tp, fn, fp, tn):
    # 查准率
    P = tp / float(tp + fp)
    # 查全率
    R = tp / float(tp + fn)
    # F1
    F1 = (2 * P * R) / (P + R)
    # print P, R, F1
    return P, R, F1

def occupancy_cal(a, b):
    result = (a + b) / float(sum_pixels)
    return result

def loss_cal(a, b):
    result = (a - b) / float(a)
    return result

print '以　二值化结果　作为　预测值，　标注框结果　作为　ground_truth　时：'.encode('gb18030')
truth, predict = occupancy_cal(white_pixels, red_pixels), occupancy_cal(white_pixels, green_pixels)
loss = loss_cal(truth, predict)
print '实际　横截面区域　占图像　{:>.3f}，　预测值为　{:>.3f}，　损失比例为　{:>.3f}'.format(truth, predict, loss).encode('gb18030')
P, R, F1 = performance(white_pixels, red_pixels, green_pixels, black_pixels)
print '查准率 P = {:>.3f},  查全率 R = {:>.3f},   F1 = {:>.3f}'.format(P, R, F1).encode('gb18030')
print
print '以　标注框结果　作为　预测值，　二值化结果　作为　ground_truth　时：'.encode('gb18030')
truth, predict = occupancy_cal(white_pixels, green_pixels), occupancy_cal(white_pixels, red_pixels)
loss = loss_cal(truth, predict)
print '实际　横截面区域　占图像　{:>.3f}，　预测值为　{:>.3f}，　损失比例为　{:>.3f}'.format(truth, predict, loss).encode('gb18030')
P, R, F1 = performance(white_pixels, green_pixels, red_pixels, black_pixels)
print '查准率 P = {:>.3f},  查全率 R = {:>.3f},   F1 = {:>.3f}'.format(P, R, F1).encode('gb18030')

以　二值化结果　作为　预测值，　标注框结果　作为　ground_truth　时：
实际　横截面区域　占图像　0.836，　预测值为　0.948，　损失比例为　-0.135
查准率 P = 0.875,  查全率 R = 0.993,   F1 = 0.930

以　标注框结果　作为　预测值，　二值化结果　作为　ground_truth　时：
实际　横截面区域　占图像　0.948，　预测值为　0.836，　损失比例为　0.119
查准率 P = 0.993,  查全率 R = 0.875,   F1 = 0.930