我希望在Python中使用图像来拟合模型(这里是2D高斯,但可能是其他东西)。
尝试使用scipy.optimize.curve_fit
我有一些问题。见下文。
让我们从一些功能开始:
import numpy as np
from scipy.optimize import curve_fit
from scipy.signal import argrelmax
import matplotlib.pyplot as plt
from matplotlib import cm
from matplotlib.patches import Circle
from tifffile import TiffFile
# 2D Gaussian model
def func(xy, x0, y0, sigma, H):
x, y = xy
A = 1 / (2 * sigma**2)
I = H * np.exp(-A * ( (x - x0)**2 + (y - y0)**2))
return I
# Generate 2D gaussian
def generate(x0, y0, sigma, H):
x = np.arange(0, max(x0, y0) * 2 + sigma, 1)
y = np.arange(0, max(x0, y0) * 2 + sigma, 1)
xx, yy = np.meshgrid(x, y)
I = func((xx, yy), x0=x0, y0=y0, sigma=sigma, H=H)
return xx, yy, I
def fit(image, with_bounds):
# Prepare fitting
x = np.arange(0, image.shape[1], 1)
y = np.arange(0, image.shape[0], 1)
xx, yy = np.meshgrid(x, y)
# Guess intial parameters
x0 = int(image.shape[0]) # Middle of the image
y0 = int(image.shape[1]) # Middle of the image
sigma = max(*image.shape) * 0.1 # 10% of the image
H = np.max(image) # Maximum value of the image
initial_guess = [x0, y0, sigma, H]
# Constraints of the parameters
if with_bounds:
lower = [0, 0, 0, 0]
upper = [image.shape[0], image.shape[1], max(*image.shape), image.max() * 2]
bounds = [lower, upper]
else:
bounds = [-np.inf, np.inf]
pred_params, uncert_cov = curve_fit(func, (xx.ravel(), yy.ravel()), image.ravel(),
p0=initial_guess, bounds=bounds)
# Get residual
predictions = func((xx, yy), *pred_params)
rms = np.sqrt(np.mean((image.ravel() - predictions.ravel())**2))
print("True params : ", true_parameters)
print("Predicted params : ", pred_params)
print("Residual : ", rms)
return pred_params
def plot(image, params):
fig, ax = plt.subplots()
ax.imshow(image, cmap=plt.cm.BrBG, interpolation='nearest', origin='lower')
ax.scatter(params[0], params[1], s=100, c="red", marker="x")
circle = Circle((params[0], params[1]), params[2], facecolor='none',
edgecolor="red", linewidth=1, alpha=0.8)
ax.add_patch(circle)
# Simulate and fit model
true_parameters = [50, 60, 10, 500]
xx, yy, image = generate(*true_parameters)
# The fit performs well without bounds
params = fit(image, with_bounds=False)
plot(image, params)
输出:
True params : [50, 60, 10, 500]
Predicted params : [ 50. 60. 10. 500.]
Residual : 0.0
现在,如果我们对边界(或约束)做同样的拟合。
# The fit is really bad with bounds
params = fit(image, with_bounds=True)
plot(image, params)
输出:
True params : [50, 60, 10, 500]
Predicted params : [ 130. 130. 0.72018729 1.44948159]
Residual : 68.1713019773
为什么在添加边界时,拟合效果不佳?
现在另一件事我不明白。当它应用于真实数据时,为什么这种拟合不稳定?见下文。
# Load some real data
image = TiffFile("../data/spot.tif").asarray()
plt.imshow(image, aspect='equal', origin='lower', interpolation="none", cmap=plt.cm.BrBG)
plt.colorbar()
# Fit is not possible without bounds
params = fit(image, with_bounds=False)
plot(image, params)
输出:
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-14-3187b53d622d> in <module>()
1 # Fit is not possible without bounds
----> 2 params = fit(image, with_bounds=False)
3 plot(image, params)
<ipython-input-11-f14c9dec72f2> in fit(image, with_bounds)
54
55 pred_params, uncert_cov = curve_fit(func, (xx.ravel(), yy.ravel()), image.ravel(),
---> 56 p0=initial_guess, bounds=bounds)
57
58 # Get residual
/home/hadim/local/conda/envs/ws/lib/python3.5/site-packages/scipy/optimize/minpack.py in curve_fit(f, xdata, ydata, p0, sigma, absolute_sigma, check_finite, bounds, method, **kwargs)
653 cost = np.sum(infodict['fvec'] ** 2)
654 if ier not in [1, 2, 3, 4]:
--> 655 raise RuntimeError("Optimal parameters not found: " + errmsg)
656 else:
657 res = least_squares(func, p0, args=args, bounds=bounds, method=method,
RuntimeError: Optimal parameters not found: Number of calls to function has reached maxfev = 1000.
和
# Fit works but is not accurate at all with bounds
params = fit(image, with_bounds=True)
plot(image, params)
输出:
True params : [50, 60, 10, 500]
Predicted params : [ 19.31770886 10.52153346 37. 1296.22524248]
Residual : 83.1944464761
答案 0 :(得分:1)
首先,您的初始参数x0
和y0
是错误的,它们不在图像的中间,但在边界处,它们应该是
x0 = int(image.shape[0])/2 # Middle of the image
y0 = int(image.shape[1])/2 # Middle of the image
将它们放在图像的边界可能会在受约束的情况下产生一些问题,因为它没有给它在某些方向上移动的空间。这是我的推测,取决于拟合方法。
同样谈论方法,curve_fit
可以使用scipy least_squares documentation中的三个中的任何一个:lm
,trf
和dogbox
:
- 'trf':信任区域反射算法,特别适用于带边界的大型稀疏问题。通常很健全的方法。
- 'dogbox':具有矩形信任区域的狗腿洞算法,典型用例是边界的小问题。不建议用于排名不足的雅可比人的问题。
- 'lm':在MINPACK中实现的Levenberg-Marquardt算法。不处理边界和稀疏的雅可比人。通常是解决小型无约束问题的最有效方法。
curve_fit
will use different methods for bounded and unbounded cases
对于无约束问题,默认为'lm',如果提供了边界,则为'trf'
所以我建议定义一个使用的方法,在更正初始参数后,我的示例使用trf
和dogbox
获得了良好的结果,但是您应该检查哪种方法更适合您的实际数据
答案 1 :(得分:1)
我写了lightweight class来做到这一点。边界没有很好地实现,但可以根据您的需要进行更改。
你在这里遇到三个主要问题:
x0
和y0
为中心的窗口中。 有两种解决问题1的方法:
median
或mode
是一个不错的选择,my class使用RANSAC
提供的blob detection algorithms算法1}}以更复杂的方式估计这一点)对于问题2.您可以使用sklearn
的{{3}}。我还写了另一个class来包装skimage
的DOG算法,以使这更容易。一旦解决问题2,问题3也会得到解决。