def my4diffEdge(img, coarseImgs, labelMap):
'''MY4
add different from Edge
'''
img = sk.color.rgb2lab(img)
# 获得4+4维 distance
degreeVectors, Ws = getVectors(img, labelMap)
weightSumVectors = getWeightSum(labelMap, degreeVectors, Ws)
diffEdges, diffNeighbors = getAllDiffEdgeAndNeighbor(labelMap, Ws)
vectors = np.append(weightSumVectors, diffEdges, 1)
# vectors = np.append(vectors,diffNeighbors,1)
vectorsTrains = []
coarseTrains = []
for coarseImg in coarseImgs:
coarseTrain, vectorsTrainTag = getCoarseTrain(coarseImg, labelMap)
vectorsTrains += list(vectors[vectorsTrainTag])
coarseTrains += list(coarseTrain)
elm = getElm(np.array(vectorsTrains), np.array(coarseTrains))
refined = elm.predict(vectors)[:, 0]
refinedImg = valueToLabelMap(labelMap, normalizing(refined))
mask = grabCut(sk.color.lab2rgb(img), refinedImg)
refinedImg = normalizing(mask * refinedImg)
return refinedImg
def precisionRecallCurve(imgGt, refinedImg, maxPoints=256):
'''
maxPoints: 返回多少个点
手写求PR线
'''
_refinedImg = normalizing(refinedImg).ravel()
_imgGt = imgGt.ravel()
l = zip(_refinedImg, _imgGt)
l.sort(key=lambda x: x[0])
n = len(_refinedImg)
step = 1. / maxPoints
P = imgGt.sum()
_l = [0] * n
summ = 0
for i in range(n):
summ += l[-i - 1][1]
_l[-i - 1] = summ
index = 0
pl, rl = [], []
for i in range(maxPoints):
thre = 0. + i * step
for j in range(index, n):
if l[j][0] >= thre:
break
index = j
TP = float(_l[j])
_P = n - index
p, r = TP / _P, TP / P
pl += [p]
rl += [r]
pl = np.array(pl)
rl = np.array(rl)
return pl, rl, None
def mergeImgs(imgName, mergeMethods):
'''
生成合并图像
'''
if isinstance(imgName, int):
imgName = IMG_NAME_LIST[imgName]
print 'img index:%d/%d' % (IMG_NAME_LIST.index(imgName),
len(IMG_NAME_LIST))
img = io.imread(IMG_DIR + imgName) / 255.
imgGt = io.imread(IMG_DIR + imgName[:-3] + 'png') / 255.
coarseDic = getCoarseDic(imgName, mergeMethods)
sumCoarseImg = sum(coarseDic.values())
sumCoarseImg = sumCoarseImg / sumCoarseImg.max()
mulImg = reduce(lambda x, y: x * y, coarseDic.values())
mulImg = normalizing(mulImg)
integratImg = integratImgsBy3way(coarseDic.values())
coarsePath = COARSE_DIR + ('%s_ME1.png' % imgName[:imgName.rindex('.')])
io.imsave(coarsePath, sumCoarseImg)
coarsePath = COARSE_DIR + ('%s_ME2.png' % imgName[:imgName.rindex('.')])
io.imsave(coarsePath, mulImg)
coarsePath = COARSE_DIR + ('%s_ME3.png' % imgName[:imgName.rindex('.')])
io.imsave(coarsePath, integratImg)
show([img, imgGt], lab=False)
print 'mergeMethods: ', ' || '.join(mergeMethods)
show(map(lambda k: coarseDic[k], mergeMethods))
print 'resoult: ', ' || '.join(['ME1', 'ME2', 'ME3'])
show([sumCoarseImg, mulImg, integratImg])
def getRefindImgsOneElmAddLabAndLbp(img, coarseImgs, labelMap):
'''MY3
use all coarse imgs to train one elm, than, predict one refinedImg
add Lab
'''
img = sk.color.rgb2lab(img)
#show([mark_boundaries(img,labelMap),imgGt])
# 获得4+4维 distance
degreeVectors, Ws = getVectors(img, labelMap)
weightSumVectors = getWeightSum(labelMap, degreeVectors, Ws)
'''add lab lbp'''
labVectors = getColorVector(img, labelMap)
lbpVectors = getLbp(img, labelMap)
vectors = np.c_[weightSumVectors, labVectors, lbpVectors]
# print 'weightSumVectors labVectors vectors.shape',weightSumVectors.shape,labVectors.shape,vectors.shape
vectorsTrains = []
coarseTrains = []
for coarseImg in coarseImgs:
coarseTrain, vectorsTrainTag = getCoarseTrain(coarseImg, labelMap)
vectorsTrains += list(vectors[vectorsTrainTag])
coarseTrains += list(coarseTrain)
elm = getElm(np.array(vectorsTrains), np.array(coarseTrains))
refined = elm.predict(vectors)[:, 0]
refinedImg = valueToLabelMap(labelMap, normalizing(refined))
return refinedImg
def getBgNoElm(img, coarseImgs=None, labelMap=None):
'''BG1
'''
img = sk.color.rgb2lab(img)
#show([mark_boundaries(img,labelMap),imgGt])
# 获得4+4维 distance
degreeVectors, Ws = getVectors(img, labelMap)
vectors = getWeightSum(labelMap, degreeVectors, Ws)
if not coarseImgs:
bg = norma(vectors2labelMap(vectors, labelMap).sum(-1))
return bg
show(bg)
pblue('befor Classifier!')
coarseImgs = [bg]
vectorsTrains = []
coarseTrains = []
for coarseImg in coarseImgs:
coarseTrain, vectorsTrainTag = getCoarseTrain(coarseImg, labelMap)
vectorsTrains += list(vectors[vectorsTrainTag])
coarseTrains += list(coarseTrain)
g()
elm = getElm(np.array(vectorsTrains), np.array(coarseTrains))
refined = elm.predict(vectors)[:, 0]
refinedImg = valueToLabelMap(labelMap, normalizing(refined))
return refinedImg
def showpr(imgName=1,
methods=["MY4", "ME1", "MEAN", "DRFI", "QCUT", "DISC2"],
lab=False):
if isinstance(imgName, int):
imgName = IMG_NAME_LIST[imgName]
img = io.imread(IMG_DIR + imgName) / 255.
imgGt = io.imread(IMG_DIR + imgName[:-3] + 'png') != 0
coarseDic = getCoarseDic(imgName, methods)
imgs = []
methods = []
for k, coarseImg in coarseDic.items():
imgs += [mark_boundaries(normalizing(coarseImg), imgGt)]
methods += [k]
print '=' * 40
print IMG_NAME_LIST.index(imgName), imgName
print 'methods:', ' || '.join(methods)
show(imgs)
show(mark_boundaries(img, imgGt))
from analysis import plotImgPr
plotImgPr(imgGt, coarseDic)
def integratImgsBy3way(refinedImgs):
def f(refinedImg):
h, w = refinedImg.shape[:2]
a, b, c, d = int(0.25 * h), int(0.75 * h), int(0.25 * w), int(0.75 * w)
center = refinedImg[a:b, c:d].sum()
ratio = float(center) / (refinedImg.sum() - center)
m, n = refinedImg.shape
distribut = 1. / mapp(
lambda x, i, j: float(x) *
((i - m / 2)**2 + (j - n / 2)**2), refinedImg, True).sum()
var = np.var(refinedImg)
return ratio, distribut, var
l = np.array(map(f, refinedImgs))
l = l / l.max(0)
l = l.sum(1)
mergeImg = sum(map(lambda x: x[0] * x[1], zip(refinedImgs, l)))
mergeImg = normalizing(mergeImg)
return mergeImg
def getRefindImgsOneElm(img, coarseImgs, labelMap):
'''MY1
use all coarse imgs to train one elm, than, predict one refinedImg
'''
img = sk.color.rgb2lab(img)
#show([mark_boundaries(img,labelMap),imgGt])
# 获得4+4维 distance
degreeVectors, Ws = getVectors(img, labelMap)
vectors = getWeightSum(labelMap, degreeVectors, Ws)
vectorsTrains = []
coarseTrains = []
for coarseImg in coarseImgs:
coarseTrain, vectorsTrainTag = getCoarseTrain(coarseImg, labelMap)
vectorsTrains += list(vectors[vectorsTrainTag])
coarseTrains += list(coarseTrain)
elm = getElm(np.array(vectorsTrains), np.array(coarseTrains))
refined = elm.predict(vectors)[:, 0]
refinedImg = valueToLabelMap(labelMap, normalizing(refined))
return refinedImg
def getRefindImgsManyElm(img, coarseImgs, labelMap):
'''MY2
for each coarse img train a elm, each elm predict a refinedImg
merge refinedImgs and normalizing to one refinedImg
'''
img = sk.color.rgb2lab(img)
#show([mark_boundaries(img,labelMap),imgGt])
# 获得4+4维 color and texture scatter degree
degreeVectors, Ws = getVectors(img, labelMap)
#
vectors = getWeightSum(labelMap, degreeVectors, Ws)
refinedImgs = []
for coarseImg in coarseImgs:
coarseTrain, vectorsTrainTag = getCoarseTrain(coarseImg, labelMap)
elm = getElm(vectors[vectorsTrainTag], coarseTrain)
refined = elm.predict(vectors)[:, 0]
refinedImg = valueToLabelMap(labelMap, normalizing(refined))
refinedImgs += [refinedImg]
# 合并 归一
refinedImgSum = integratImgsBy3way(refinedImgs)
return refinedImgSum
